The voice AI agents market is projected to grow from $2.4 billion in 2024 to $47.5 billion by 2034, a 34.8% compound annual growth rate. Yet only 1% of enterprises consider their AI deployments “mature” and fewer than 10% of AI use cases make it past pilot stage.

The models work but the gap is in how these systems are designed for real human interaction in enterprise collaboration, where voice commands trigger workflows, meetings have audiences and mistakes carry social weight. This article is about where they live and how to solve them.

Where enterprise voice AI breaks down

81% of consumers now use voice technology daily or weekly, but satisfaction hasn’t kept up. 65% of voice assistant users report regular misunderstandings. 41% admit to yelling at their voice assistant when things go wrong. These same people walk into work the next morning and are expected to trust a voice agent with their calendar, their meetings and their messages. The frustration they’ve learned at home sets the baseline expectation at work.

Most teams look at numbers like these and reach for technical fixes: Better speech recognition models, lower Word Error Rate (WER), faster processing. But WER tells you how well your system transcribed audio. It says nothing about whether someone trusted the agent enough to use it in front of their manager, or whether they’ll open it again next week. In enterprise collaboration, one misunderstood instruction and someone has a calendar invite they never asked for.

The root of the problem is a design assumption that keeps getting repeated: Treating voice AI as text with a microphone attached. Voice has its own constraints. Anything beyond a 500ms response breaks conversational flow. Commands arrive mixed in with meeting crosstalk and open-office noise. Users can’t scroll back through what the agent said. And when the system gets something wrong in a meeting, the embarrassment lands differently than a typo in a chat window.

When you map user journeys for voice-driven enterprise workflows, the breakdowns don’t cluster around transcription failures. They cluster around moments of social risk: Issuing a command in front of an executive, trusting the system to send the right message or waiting in awkward silence while the agent processes. Nielsen’s usability heuristics help explain why. Visibility of system status means something entirely different in a voice-only interface where there’s no progress bar, no loading spinner. Users are left interpreting silence, and that ambiguity is one of the strongest predictors of early abandonment.

UX principles for building voice AI agents

There’s a reason conversations have rhythm. Sacks, Schegloff and Jefferson (1974) documented that people take turns in speech on roughly 200-ms cycles, regardless of language. When a voice agent takes even slightly longer than that, the interaction starts to feel off. People won’t say ‘the latency was too high’. They’ll say the thing felt clunky, or they’ll just stop using it.

This means agents need to acknowledge while processing. ‘Got it, looking that up..’ feels collaborative. People describe faster-responding systems as “more helpful” even when task completion rates are identical. Google’s Speech-to-Text documentation recommends 100-ms frame sizes for streaming applications. Dan Saffer’s work on microinteractions is useful here. Think about what makes a phone call feel natural: The ‘mm-hmm’ that says someone is listening, a pause before an answer, the rising voice inviting you to keep going. Voice agents need all of that. None of it shows up in a spec, but it separates a system people tolerate from one they want to use.

Recovery matters as much as performance. People are forgiving the first time a voice agent gets something wrong. Second time, doubt creeps in. By the third, they’ve filed it under “doesn’t work” – thus impacting trust. The agent needs to explicitly state when it is confused or when it cannot give the correct response and offer workarounds like closest reference documents or next steps to create trust and transparency.

Implicit confirmation is another principle that pays off immediately in enterprise settings. ‘I’ve sent an updated sales invoice to your inbox’ works better than ‘Did you send a sales invoice to me? Please say yes or no’. There’s a half-second pause right before someone issues a voice command where they’re doubting if the agent is going to give the right response and if they should proceed. Good confirmation design takes that social risk down.

Finally, the environment is a design constraint, not a testing variable. Open offices, conference rooms, mobile use in transit, hybrid meetings: Each sounds different, and each creates different failure modes. Denoising and automatic speaker diarization aren’t nice-to-have features. They are table stakes.

The UX research playbook for building effective voice AI agents

Standard usability testing assumes the interface is visible and the system behaves the same way every time. Voice AI agents break both of those assumptions. The system’s behavior is non-deterministic, the interaction leaves no visual trace and the environment changes everything. The research approach has to account for all of that.

Contextual inquiry is essential because the acoustic environment is the primary design constraint. Observing someone use a voice agent while a coworker’s speakerphone bleeds through a conference room wall tells you more about what needs to change than any controlled study can. Think-aloud protocols need adaptation here too. Participants are already talking to the system, so concurrent think-aloud creates interference. The workaround is retrospective think-aloud with recordings, letting participants replay interactions and narrate what they were thinking at each point.

Field research only captures a snapshot, though. Diary studies take on a different role with AI voice agents than with traditional software. Instead of tracking feature usage, they track trust over time. Participants log not just what happened, but whether they’d repeat the interaction in front of colleagues. That’s how you spot trust starting to slip before your retention numbers do. Experience sampling picks up what even diary studies miss: You check in with people at random points while they’re actually using the agent, not after. Ask someone in a debrief and they’ll tell you it was fine. Their notes from the moment tell a different story.

Then there is Quantitative UX Research and Behavioral Data Collection. Look at conversation logs: How often does the agent fall back to a generic response? Where do people abandon a request halfway through? Which user segments hit more errors than others? That data shows you where the system is failing at scale. Pairing this with qualitative findings turns isolated observations into product decisions.

But the numbers that matter most aren’t the obvious ones. The pattern that keeps showing up is how often task completion and user satisfaction tell completely different stories. Someone finishes a task and still walks away frustrated: ‘It worked but I wouldn’t do that again in a meeting’. You only catch that divergence by pairing something like the System Usability Scale with behavioral data and qualitative follow-ups. Measurement works best when you’re looking at multiple levels at once. At the conversation level, you care about how the agent handles interruptions and how often it hits a fallback. At the business level, the question is simple: Did people keep using it after the first week? The interesting stuff lives in the gaps between those levels, and you’ll only see it if research teams are involved from the beginning, not called in after the product decisions are already locked.

Testing across the full range of speech patterns, accents and accessibility needs the product will encounter in production also reshapes product direction in ways teams don’t expect. The Speech Accessibility Project, run by the University of Illinois with Google, Apple and Amazon, trained models on a broader set of speech samples and saw accuracy jump by 18 to 60% for non-standard speech patterns. Card sorting exercises with diverse user groups regularly upend what product teams assumed users wanted. Also, curb-cut effects are real in voice AI: Building for users who depend entirely on voice produces better experiences across the board.

How UX research shapes agentic voice AI

When a voice agent moves from executing single commands to acting autonomously across enterprise workflows, the UX research problem changes. ‘Prepare tomorrow’s client meeting’ might involve pulling calendar data, finding documents and writing up a summary. Zoom’s AI Companion 3.0 works this way. The research question is no longer ‘did the system understand the words?’ It’s ‘does the person trust what the agent did on their behalf?’

The trust problem comes down to mental models. If someone says ‘reschedule tomorrow’s meetings’, they’re picturing the whole job: Check for conflicts, move the time slots, update the invites, notify the attendees. If the agent only moves the slots and silently drops the rest, that half-finished job feels worse than if it had just said ‘I can’t do that’. People shrug off an honest limitation. They don’t shrug off finding out an hour later nobody got notified.

What makes enterprise different is that the agent’s actions affect other people. An enterprise voice agent that misfires wastes your colleague’s time, sends your manager the wrong information or derails a meeting you weren’t even in. When the agent gets it wrong, other people pay the price and that makes people far less forgiving. A good way to catch these problems early in research is to ask participants to walk through what they expect the agent to do before it does it, then compare that against what actually happens. Those mismatches are early warnings. They’ll show up in your research months before they show up in support tickets or churn.

‘Least surprise’ carries extra weight in agentic contexts. Even when multiple things are happening behind the scenes, the person should get back one clear answer. Giving feedback during wait times, even “Let me pull together a few things for that,” buys the system a few seconds without silence. Journey mapping shows users lose confidence in the middle of a request, during that gap. That’s the moment to get right.

Teams also need to plan for novelty wearing off. Early on, people give the system a pass when it stumbles. That wears off fast. Around week two or three, the comparison shifts. People stop thinking ‘that’s pretty good for AI’ and start thinking ‘my admin assistant would have gotten that right’. At work, everyone already knows what competent help looks like: The assistant who juggles calendars, the IT person who fixes things without being asked twice, the colleague who never forgets to send the agenda. That’s the bar, and the only way to see whether the system is going to clear it over time is longitudinal research.

Design problems, not engineering ones

The problems with enterprise voice AI aren’t technical mysteries. The models work. What’s been missing is treating voice AI as a UX problem from the start, applying research practice to the specific challenges that voice and agentic AI create in enterprise collaboration. Social risk, autonomous trust decisions, the gap between what the system can do and what people will actually rely on: These are design problems, not engineering ones.

As voice AI agents grow more autonomous, the question researchers and builders should be asking together isn’t ‘does this work?’ It’s ‘do people trust it enough to let it act on their behalf, in front of other people, without checking its work first?’ That’s the real adoption threshold. The methods and principles to get there are well understood. What matters now is whether teams put UX researchers in the room early enough to use them.

Disclaimer: The views expressed in this article are my own and do not represent those of my employer.

This article is published as part of the Foundry Expert Contributor Network.
Want to join?

Artificial intelligence and related technologies are evolving rapidly, but until recently, Java developers had few options for integrating AI capabilities directly into Spring-based applications. Spring AI changes that by leveraging familiar Spring conventions such as dependency injection and the configuration-first philosophy in a modern AI development framework.

My last tutorial demonstrated how to configure Spring AI to use a large language model (LLM) to send questions and receive answers. While this can be very useful, it does not unlock all the power that AI agents provide. In this article, you will learn exactly what an agent is and how to build one manually, then you’ll see how to leverage Spring AI’s advanced capabilities and support for building robust agents using familiar Spring conventions.

What is an AI agent?

Before we dive into building an AI agent, let’s review what an agent actually is. While standard LLM interactions consist of sending a request and receiving a response, an agent is more than a chatbot and follows a more complicated set of tasks. An AI agent typically performs the following steps in sequence. We call this sequence the agent loop:

• Receives a goal
• Interprets the user’s intent
• Plans actions
• Selects tools
• Executes tools
• Observes results
• Refines strategy
• Iterates the process
• Produces a final answer
• Terminates safely

In essence, an agent accepts a user request, uses an LLM to interpret what the user really wants, and decides if it can respond directly or if it needs external support. Once a request is accepted, the agent chooses the tools it will use from the set provided, calls tools for any information it needs, and receives and incorporates that output into its working context. Next, it decides whether the preliminary result is sufficient or if it needs to call additional tools to reach a satisfactory end. The agent repeats this plan-act-observe cycle until the objective is satisfied. Once satisfied, it returns a completed answer. It stops execution based on a completion indicator, safety checks, or the given iteration limit.

The following diagram visualizes the agent loop:

Steven Haines

If this sounds a little abstract, try asking your favorite chatbot, such as ChatGPT, to help you do something that requires a knowledge base and a few steps. In the example below, I prompted ChatGPT to help me bake a cake:

I want to bake a cake. Can you tell me what to do step-by-step, one step at a time? Tell me each step to perform and I will tell you the results. Please start with the first step.

The model in this case responded with a list of ingredients, then asked if I had everything I needed. I responded that I did not have eggs, so it offered a list of substitutions. Once I had all the ingredients, the model told me to mix them and continued with step-by-step instructions to bake a cake. As a test, once the cake was baking, I reported that I thought it might be burning. The model responded that I should turn down the oven temperature, cover the cake with aluminum foil, and describe what it looked like to determine if it could be salvaged.

So, in this exercise, the LLM planned out what to do, walked through the process one step at a time, and used me as a “tool” to perform the actions needed and report the results. When things didn’t work out as expected, such as missing ingredients or a burning cake, it adapted the plan to still achieve its objective. This is exactly what agents do, but relying on a set of programmatic tools, rather than a hungry human, to perform the needed actions. This may be a silly example, but it illustrates the key elements of agent behavior, including planning, use of tools, and the ability to adapt to changing circumstances.

As another example, consider the difference between using a ChatGPT conversation to generate code versus using an AI coding tool like Claude. ChatGPT responds to your prompts with code to copy-and-paste into your application. It is up to you to paste in the code, and also build and test it. Claude, on the other hand, has its own tools and processes. Namely, it can search through the files on your file system, create new files, run build scripts like Maven, see the results, and fix build errors. Whereas ChatGPT is a chatbot that relies on you to do the work, Claude is a complete coding agent: You provide it with an objective and it does the coding for you.

Also see: What I learned using Claude Sonnet to migrate Python to Rust.

Building a Spring AI agent

Now that you have a sense of what an AI agent is, let’s build one with Spring AI. We’ll do this in two phases: First, we’ll build our own agent loop and do everything manually, so that you can understand exactly how agents work and what Spring AI does behind the scenes; then we’ll leverage the capabilities built into Spring AI to make our job easier.

For our example, we’ll build the product search agent illustrated in the diagram below:

Steven Haines

Note that this demonstration assumes you are familiar with Java development and with Spring coding conventions.

Defining the product search tool

To start, we have a database that contains over 100 products and a Spring MVC controller to which we can POST a natural language query for products. As an example, we might enter, “I want sports shoes that cost under $120.” The controller calls a service that leverages our product search agent to work with an LLM and searches the database. The tool that we’re building uses a repository that has a simple keyword search query that runs against product names and descriptions. The LLM is responsible for determining the user’s intent, choosing the most applicable keywords to search for, calling the tool to retrieve products that match each keyword, and returning the list of relevant products.

Here’s the Product class:

package com.infoworld.springagentdemo.model;

import jakarta.persistence.Entity;
import jakarta.persistence.GeneratedValue;
import jakarta.persistence.GenerationType;
import jakarta.persistence.Id;

@Entity
public class Product {
    @Id
    @GeneratedValue(strategy = GenerationType.AUTO)
    private Long id;
    private String name;
    private String description;
    private String category;
    private Float price;

    public Long getId() {
        return id;
    }
    public void setId(Long id) {
        this.id = id;
    }
    public String getName() {
        return name;
    }
    public void setName(String name) {
        this.name = name;
    }
    public String getDescription() {
        return description;
    }
    public void setDescription(String description) {
        this.description = description;
    }
    public String getCategory() {
        return category;
    }
    public void setCategory(String category) {
        this.category = category;
    }
    public Float getPrice() {
        return price;
    }
    public void setPrice(Float price) {
        this.price = price;
    }
}

The Product class is a JPA entity with an id, name, description, category, and price. The repository is a JpaRepository that manages products:

package com.infoworld.springagentdemo.repository;

import java.util.List;

import com.infoworld.springagentdemo.model.Product;

import org.springframework.data.jpa.repository.JpaRepository;
import org.springframework.data.jpa.repository.Query;
import org.springframework.data.repository.query.Param;

public interface ProductRepository extends JpaRepository {
    @Query("""
       SELECT p FROM Product p
       WHERE lower(p.name) LIKE lower(concat('%', :query, '%'))
          OR lower(p.description) LIKE lower(concat('%', :query, '%'))
    """)
    List search(@Param("query") String query);
}

We added a custom search method with a query that returns all products with a name or description that matches the specified query string.

Now let’s look at the ProductSearchTools class:

package com.infoworld.springagentdemo.ai.tools;

import java.util.List;

import com.infoworld.springagentdemo.model.Product;
import com.infoworld.springagentdemo.repository.ProductRepository;

import org.springframework.ai.tool.annotation.Tool;
import org.springframework.stereotype.Component;

@Component
public class ProductSearchTools {

    private final ProductRepository repository;

    ProductSearchTools(ProductRepository repository) {
        this.repository = repository;
    }

    @Tool(description = "Search products by keyword")
    public List searchProducts(String keyword) {
        return repository.search(keyword);
    }
}

The ProductSearchTools class is a Spring-managed bean, annotated with the @Component annotation, and defines a searchProducts() method that calls the repository’s search() method. You’ll learn more about the @Tool annotation when we use Spring AI’s built-in support for tools. For now, just note that this annotation marks a method as a tool that the LLM can call.

Developing the agent

With the tool defined, let’s look at the ManualProductSearchAgent, which is the explicit version of our search agent in which we define our agent loop manually:

package com.infoworld.springagentdemo.ai.agent;

import java.util.ArrayList;
import java.util.List;

import com.fasterxml.jackson.core.JsonProcessingException;
import com.fasterxml.jackson.databind.ObjectMapper;
import com.infoworld.springagentdemo.ai.tools.ProductSearchTools;
import com.infoworld.springagentdemo.model.Product;

import org.springframework.ai.chat.client.ChatClient;
import org.springframework.ai.chat.messages.AssistantMessage;
import org.springframework.ai.chat.messages.Message;
import org.springframework.ai.chat.messages.SystemMessage;
import org.springframework.ai.chat.messages.UserMessage;
import org.springframework.ai.chat.prompt.Prompt;
import org.springframework.stereotype.Component;

@Component
public class ManualProductSearchAgent {

    private final ChatClient chatClient;
    private final ProductSearchTools productSearchTools;
    private final ObjectMapper objectMapper;
    private final int MAX_ITERATIONS = 10;

    public ManualProductSearchAgent(ChatClient.Builder chatClientBuilder,
                                    ProductSearchTools productSearchTools) {
        this.chatClient = chatClientBuilder.build();
        this.productSearchTools = productSearchTools;
        this.objectMapper = new ObjectMapper();
    }

    public List search(String userInput) {

        List messages = new ArrayList();

        // System Prompt with Tool Specification
        messages.add(new SystemMessage("""
                You are a product search agent.
                
                You have access to the following tool:
                
                Tool Name: searchProducts
                Description: Search products by keyword
                Parameters:
                {
                  "keyword": "string"
                }
                
                You may call this tool multiple times to refine your search.
                
                If the user request is vague, make reasonable assumptions.
                
                If the user asks about products in a certain price range, first search for the products and then filter
                the results based on the price. Each product is defined with a price.
                
                You must respond ONLY in valid JSON using one of these formats:
                
                To call a tool:
                {
                  "action": "tool",
                  "toolName": "searchProducts",
                  "arguments": {
                    "keyword": "..."
                  }
                }
                
                When finished:
                {
                  "action": "done",
                  "answer": "final response text",
                  "products": "a list of matching products"
                }
                
                Do not return conversational text.
                """));

        messages.add(new UserMessage(userInput));

        // Manual Agent Loop
        int iteration = 0;
        while (iteration++  result =
                                productSearchTools.searchProducts(keyword);

                        String observation =
                                objectMapper.writeValueAsString(result);

                        // Feed Observation Back Into Context
                        messages.add(new AssistantMessage(response));

                        messages.add(new SystemMessage("""
                                Tool result from searchProducts:
                                """ + observation));
                    }
                }
            } catch (JsonProcessingException e) {
                System.out.println(e.getMessage());
            }
        }
        return new ArrayList();
    }
}

The ManualProductSearchAgent constructor accepts a ChatClient.Builder that it uses to build a ChatClient. If you have not read the Getting Started with Spring AI article yet, the ChatClient class is Spring AI’s abstraction to interacting with an LLM. It is configured in the application.yaml file as follows:

spring:
  application:
    name: spring-aiagent-demo
  ai:
    openai:
      api-key: ${OPENAI_API_KEY}
      chat:
        options:
          model: gpt-5
          temperature: 1
  jpa:
    defer-datasource-initialization: true

In this case, I opted to use OpenAI and pass in my API key as an environment variable. It uses the gpt-5 model with a temperature of 1, which is required by Spring AI. (See the first tutorial if you need to more information.) If you download the source code and define an OPEN_API_KEY environment variable, you should be able to run the code.

Next, the constructor accepts a ProductSearchTools instance and then creates a Jackson ObjectMapper to deserialize JSON into Java classes. The search() method is where the agent is defined. First, it maintains a list of messages that will be sent to the LLM. These come in three forms:

• SystemMessage: The message that defines the role of the agent. It defines the steps it should take, as well as the rules it should follow.
• UserMessage: The message that the user passed in, such as “I want sports shoes that cost less than $120.”
• AssistantMessage: These messages contain the history of the conversation so that the LLM can follow the conversation.

The above prompt defines the initial system message. We inform the LLM that it is a product search agent that has access to one tool: the searchProducts tool. We provide a description of the tool and tell the LLM that it must pass a keyword parameter as a String. Next, we tell it that it can call the tool multiple times and give it some additional instructions. I purposely added the instruction that if the user asks for products in a certain price range, the LLM should first search for the products and then filter on the price. Before I added this instruction, the LLM included the price in the search, which yielded no results. The key takeaway here is that you are going to need to experiment with your prompt to get the results you are seeking.

Next, we tell the LLM that, to call a tool, it should return an action of “tool” and a tool name and arguments. If we gave it more tools, it is important that it tells us exactly what tool to execute. Finally, we define the format of the message it should return when it is finished; namely, an action of “done,” an answer String, and a list of products.

After adding our prompt as a SystemMessage, we add the user’s query as a UserMessage. Now, the LLM knows what it is supposed to do, what tools it has access to, and the goal that it must accomplish.

Implementing the agent loop

Next, we implement our agent loop. We defined a MAX_ITERATIONS constant of 10, which means that we will only call the LLM a maximum of 10 times. The number of iterations you need in your agent will depend on what you are trying to accomplish, but the purpose is to restrict the total number of LLM calls. You would not want it to get into an infinite loop and consume all your API tokens.

The first thing we do in our agent loop is construct a prompt from our list of messages and call the LLM. The content() method returns the LLM response as a String. We could have used the entity() method to convert the response to an AgentDecision class instance, but we leave it as a String and manually convert it using Jackson so that we can add the response as an AssistantMessage later to keep track of the conversation history. An AgentDecision is defined as follows:

package com.infoworld.springagentdemo.ai.agent;

import java.util.List;
import java.util.Map;

import com.infoworld.springagentdemo.model.Product;

public record AgentDecision(
    String action,
    String toolName,
    Map arguments,
    String answer,
    List products) {
}

We check the AgentDecision action to see if it is “done” or if it wants to invoke a “tool.” If it is done, then we return the list of products that it found. If it wants to invoke a tool, then we check the tool that it wants to invoke against the name “searchProducts,” extract the keyword argument that it wants to search for, and call the ProductSearchTool’s searchProducts() method. We save the query response and add it as a new SystemMessage and we store the LLM’s request for the tool call as an AssistantMessage.

We continue the process until we reach the maximum number of iterations or the LLM reports that it is done.

Testing the AI agent

You can use the following controller to test the agent:

package com.infoworld.springagentdemo.web;

import java.util.List;

import com.infoworld.springagentdemo.model.Product;
import com.infoworld.springagentdemo.model.SearchRequest;
import com.infoworld.springagentdemo.service.ProductService;

import org.springframework.web.bind.annotation.GetMapping;
import org.springframework.web.bind.annotation.PostMapping;
import org.springframework.web.bind.annotation.RequestBody;
import org.springframework.web.bind.annotation.RestController;

@RestController
public class ProductController {
    private ProductService productService;

    public ProductController(ProductService productService) {
        this.productService = productService;
    }

    @GetMapping("/products")
    public List getProducts() {
        return productService.findAll();
    }

    @PostMapping("/search")
    public List searchProducts(@RequestBody SearchRequest request) {
        return productService.findProducts(request.query());
    }

    @PostMapping("/manualsearch")
    public List manualSearchProducts(@RequestBody SearchRequest request) {
        return productService.findProductsManual(request.query());
    }
}

This controller has a getProducts() method that returns all products, a searchProducts() method that will use Spring AI’s built-in support for tools, and a manualSearchProducts() method that calls the agent we just built. The SearchRequest is a simple Java record and is defined as follows:

package com.infoworld.springagentdemo.model;

public record SearchRequest(String query) {
}

The ProductService is a passthrough service that invokes the agent, or the repository in the case of listing all products:

package com.infoworld.springagentdemo.service;

import java.util.List;

import com.infoworld.springagentdemo.ai.agent.ManualProductSearchAgent;
import com.infoworld.springagentdemo.ai.agent.ProductSearchAgent;
import com.infoworld.springagentdemo.model.Product;
import com.infoworld.springagentdemo.repository.ProductRepository;

import org.springframework.stereotype.Service;

@Service
public class ProductService {
    private final ProductRepository productRepository;
    private final ProductSearchAgent productSearchAgent;
    private final ManualProductSearchAgent manualProductSearchAgent;

    public ProductService(ProductRepository productRepository, ProductSearchAgent productSearchAgent, ManualProductSearchAgent manualProductSearchAgent) {
        this.productRepository = productRepository;
        this.productSearchAgent = productSearchAgent;
        this.manualProductSearchAgent = manualProductSearchAgent;
    }

    public List findAll() {
        return productRepository.findAll();
    }

    public List findProducts(String query) {
        return productSearchAgent.run(query);
    }

    public List findProductsManual(String query) {
        return manualProductSearchAgent.search(query);
    }
}

You can test the application by POSTing a request to /manualsearch with the following body:

{
    "query": "I want sports shoes under $120"
}

Your results may be different from mine, but I saw the LLM searching for the following keywords:

Searching products by keyword: sports shoes
Searching products by keyword: running shoes
Searching products by keyword: sports shoes
Searching products by keyword: running shoes
Searching products by keyword: athletic shoes

And I received the following response:

[
    {
        "category": "Clothing",
        "description": "Lightweight mesh running sneakers",
        "id": 24,
        "name": "Running Shoes",
        "price": 109.99
    },
    {
        "category": "Clothing",
        "description": "Cross-training athletic shoes",
        "id": 83,
        "name": "Training Shoes",
        "price": 109.99
    }
]

So, the agent effectively determined what I meant by “sports shoes,” selected some relevant keywords to search for, filtered the products based on price, and returned a list of two options for me. Because LLMs are not deterministic, your results may be different from mine. For example, in other runs with the same query, the agent searched for different keywords and returned a larger list. But being able to translate a natural language query into a set of database queries and find relevant results is impressive!

Spring AI’s built-in support for developing agents

Now that you understand what an agent loop is, what it does, and how to handle tool executions, let’s look at Spring AI’s built-in support for managing its own agent loop and tool execution. Here is our updated ProductSearchAgent code:

package com.infoworld.springagentdemo.ai.agent;

import java.util.ArrayList;
import java.util.List;

import com.infoworld.springagentdemo.ai.tools.ProductSearchTools;
import com.infoworld.springagentdemo.model.Product;

import org.springframework.ai.chat.client.ChatClient;
import org.springframework.ai.chat.messages.Message;
import org.springframework.ai.chat.messages.SystemMessage;
import org.springframework.ai.chat.messages.UserMessage;
import org.springframework.ai.chat.prompt.Prompt;
import org.springframework.ai.tool.method.MethodToolCallbackProvider;
import org.springframework.stereotype.Component;

@Component
public class ProductSearchAgent {

    private final ChatClient chatClient;
    private final ProductSearchTools productSearchTools;

    public ProductSearchAgent(ChatClient.Builder chatClientBuilder, ProductSearchTools productSearchTools) {
        this.chatClient =  chatClientBuilder.build();
        this.productSearchTools = productSearchTools;
    }

    public List run(String userRequest) {

        Prompt prompt = buildPrompt(userRequest);

        AgentResponse response = chatClient
                .prompt(prompt)
                .toolCallbacks(
                        MethodToolCallbackProvider.builder().toolObjects(productSearchTools).build()
                )
                .call()
                .entity(AgentResponse.class);

        System.out.println(response.answer());
        return response.products();
    }

    private Prompt buildPrompt(String userRequest) {

        List messages = new ArrayList();

        // 1. System message: defines the agent
        messages.add(new SystemMessage("""
You are a product search agent.

Your responsibility is to help users find relevant products using the available tools.

Guidelines:
- Use the provided tools whenever product data is required.
- You may call tools multiple times to refine or expand the search.
- If the request is vague, make reasonable assumptions and attempt a search.
- Do not ask follow-up questions.
- Continue using tools until you are confident you have the best possible results.

If the user asks about products in a certain price range, first search for the products and then filter
the results based on the price. Each product is defined with a price.

When you have completed the search process, return a structured JSON response in this format:

{
  "answer": "...",
  "products": [...]
}

Do not return conversational text.
Return only valid JSON.
"""));

        // Add the user's request
        messages.add(new UserMessage(userRequest));

        return new Prompt(messages);
    }
}

As I mentioned earlier, the ProductSearchTools’ searchProducts() method is annotated with the @Tool annotation. This annotation has special meaning for Spring AI if we add a toolCallbacks() method call to our LLM call. In this case, we autowire the ProductSearchTools into our constructor and then invoke the toolCallbacks() method in our LLM call, passing it a list of all the classes containing tools we want to give the LLM access to in a MethodToolCallbackProvider.builder().toolObjects() call. Spring AI will see this list of tools and do a few things:

1. Introspect all methods annotated with the @Tool annotation in the provided classes.
2. Build the tool specification and pass it to the LLM for us, including the description of the tool and the method signature, which means that we no longer need to explicitly define the tool specification in our SystemPrompt.
3. Because it has access to call the tools, the ChatClient’s call() method will run in its own agent loop and invoke the tools it needs for us.

Therefore, the response we receive will be the final response from the LLM with our list of products, so we do not need to build an agent loop ourselves. We build our prompt with a system prompt (which again does not have the tool specification) and the user’s request. We then make a single call to the call() method, which performs all the actions it needs to arrive at a conclusion.

You can test it by executing a POST request to /search with the same SearchRequest payload and you should see similar results. Claude was kind enough to generate my test products for me, so feel free to search for shirts, jackets, pants, shoes, and boots. You can find the full list of products preconfigured in the database in the src/resources/import.sql file.

Conclusion

This tutorial introduced you to using Spring AI to build AI agents. We began by reviewing what an agent is, which in its simplest form is a class that receives an objective. The agent makes repeated calls to an LLM, first to make a step-by-step plan to meet the objective, and then to execute the plan using whatever tools were provided.

To give you a really good sense of what agents are, we manually built an agent loop, executed tools, and interacted with the LLM through SystemMessages, AssistantMessages, and UserMessages. Then, we leveraged Spring AI’s capabilities to let the agent execute tools on its own. Spring AI provides Spring developers with all the tools needed to build complex AI applications, including an LLM abstraction, through the ChatClient class and a YAML configuration, built-in support for discovering and executing tools, and a built-in agent loop to remove the complexity of manually writing code yourself.

With what you learned in this tutorial, you should be able to start building Spring AI agents on your own. You could try developing your own coding assistant, an agent that downloads and summarizes articles from the Internet, or even an agent that translates natural language into database queries. All you need to do is build the tools, write the prompts, and leverage Spring AI’s agent development capabilities and support.

Computer vision teams face an uncomfortable reality. Even as annotation costs continue to rise, research consistently shows that teams annotate far more data than they actually need. Sometimes teams annotate the wrong data entirely, contributing little to model improvements. In fact, by some estimates, 95% of data annotations go to waste.

The problem extends beyond cost. As I explored in my previous article on annotation quality, error rates average 10% in production machine learning (ML) applications. But there’s a deeper issue that precedes annotation quality: Most teams never develop systematic approaches to selecting which data needs annotation in the first place. This is largely because annotation often remains siloed from data curation and model evaluation, making it impossible to act on the full picture.

Safety-critical models, such as models for autonomous vehicles (AV) with multi-sensor perception stacks, require highly accurate 2D bounding boxes and 3D cuboid annotations. Without intelligent data selection, teams find themselves not only collecting vast amounts of data but also labeling millions of redundant samples while missing the edge cases that actually improve model performance.

When tools become barriers

The conventional approach treats annotation as an isolated workflow: Collect data, export to a labeling platform, wait for humans to label data, import labels, discover problems, go back to the annotation vendor, and repeat. This fragmentation creates two critical gaps that turn annotation into a development bottleneck rather than an enabling capability.

No systematic data selection

Random sampling and “label everything” approaches waste annotation budgets on redundant samples. Teams annotating AV datasets might label 100,000 highway cruise images that provide minimal new information while missing rare scenarios like emergency vehicle encounters or unusual weather conditions.

Lost context across tool boundaries

When annotation lives in one platform, curation in another, and model evaluation in a third, teams lose critical context at each handoff. Data scientists spend 80% of their time curating data, yet most of this effort happens in ad hoc, disconnected ways that don’t inform downstream annotation decisions.

Some estimates indicate that ~45% of companies now use four or more tools simultaneously, cobbling together partial solutions that impact budgets and timelines.

Curate first: A paradigm shift in ML workflows

The “curate first, then annotate” approach inverts the conventional wisdom. Instead of treating data curation as a second step in development, curation becomes the foundation that drives intelligent annotation decisions. This methodology recognizes that annotation isn’t primarily a labeling problem—it’s a data understanding problem.

Strategic data selection focuses on annotation where it matters

Zero-shot coreset selection represents a breakthrough in pre-annotation intelligence. Using pre-trained foundation models to analyze unlabeled data, this technique scores each sample based on unique information contribution, automatically filtering redundant examples.

The methodology works through iterative subspace sampling:

1. Embedding computation: Foundation models generate high-dimensional representations capturing semantic content.
2. Uniqueness scoring: Each sample receives a score indicating information diversity relative to existing selections.
3. Iterative selection: Samples with the highest uniqueness scores enter the training set.
4. Redundancy elimination: Visually similar samples get deprioritized automatically.

Benchmarks on ImageNet demonstrate that this approach achieves the same model accuracy with just 10% of training data, eliminating annotation costs for over 1.15 million images.

Voxel51

To put it in perspective, for a 100,000-image dataset at typical rates of $0.05 to $0.09 per object, strategic selection can save ~$81K in annotation costs while improving model generalization on edge cases.

Programmatically:

import fiftyone.zoo as foz
from zcore import zcore_scores, select_coreset

dataset = foz.load_zoo_dataset("quickstart")
model = foz.load_zoo_model("clip-vit-base32-torch")
embeddings = dataset.compute_embeddings(model, batch_size=2)

scores = zcore_scores(embeddings, use_multiprocessing=True, num_workers=4)
coreset = select_coreset(dataset, scores, coreset_size=int(0.1 * len(dataset)))

Embedding-based curation

This approach surfaces the samples that will contribute most to model learning, transforming annotation from a volume game into a strategic exercise.

Modern platforms enable embedding-based curation through straightforward workflows. For example, you can leverage computed embeddings to identify the most unique samples in the embedding space using a k-nearest-neighbors calculation. Those samples are then prioritized for annotation.

import fiftyone as fo
import fiftyone.brain as fob
import fiftyone.zoo as foz

# Load your unlabeled dataset
dataset = fo.Dataset.from_dir(
    dataset_dir="/path/to/images",
    dataset_type=fo.types.ImageDirectory,
)

# Generate embeddings using pre-trained model
model = foz.load_zoo_model("clip-vit-base32-torch")
dataset.compute_embeddings(model, embeddings_field="embeddings")

# Perform uniqueness-based selection
fob.compute_uniqueness(dataset, embeddings_field="embeddings")

# Sort by uniqueness score to prioritize diverse samples
unique_view = dataset.sort_by("uniqueness", reverse=True)

# Select top 10% most informative samples for annotation
samples_to_annotate = unique_view.take(len(dataset) // 10)

Voxel51

Model analysis results feed into prioritizing what to label

Once you have trained a baseline model on your initial curated subset, you can shift from pure data exploration to targeted improvement. Instead of randomly selecting the next batch, use the model’s own predictions to identify “hard” samples where the model is confused or uncertain.

The most effective workflow intersects uncertainty with uniqueness. This ensures you prioritize valid edge cases that drive better model understanding, rather than just noise (for example, blurry images which are inherently low-confidence).

We can filter programmatically for this “Goldilocks zone” of high uniqueness and low confidence.

from fiftyone import ViewField as F

# Filter for samples where model confidence is low
hard_samples = dataset.match(F("predictions.confidence") 

Quantifying the curation advantage

The financial impact of curation-first workflows manifests across multiple dimensions, with organizations reporting cost and efficiency improvements.

• Reduced annotation volume: Curation achieves equivalent model performance with 60% to 80% less annotated data.
• Lower error correction costs: Finding and fixing labeling mistakes early reduces expensive rework cycles that typically add 20% to 40% to project budgets.
• Minimized tool licensing and coordination overhead: Unified workflows eliminate redundant platform costs that average $50K annually per tool and minimize handoffs.
• Faster iteration cycles: Targeted annotation and validation eliminate weeks of review cycles.

A mid-sized AV team annotating 500K samples monthly at $0.07 per object can reduce this from $35K to $14K through intelligent selection, leading to an annual savings of ~$336K.

Impact on development teams: From reactive to strategic

The shift to curation-first methodologies fundamentally changes how ML engineering teams operate, moving them from reactive problem-solving to proactive dataset optimization.

Workflow transformation

Traditional workflow:

Data collection → Data annotation → Model training → Discover failures → Debug → Reannotate → Retrain

Curation-first workflow:

Data collection → Intelligent curation → Targeted annotation → Continuous validation → Model training → Strategic expansion

This reordering frontloads data understanding, helping identify issues when they’re cheapest to fix. Teams report improvements in doing real work as engineers shift their focus from tedious quality firefighting to strategic model improvement.

Best practices: Implementing curation-driven annotation

Successful implementations follow established patterns that balance automation with human expertise.

Start with embedding-based exploration

Before annotating anything, generate embeddings and visualize your dataset’s distribution. This reveals the structure and distribution of your dataset. For example, tight clusters indicate redundancy, or sparse regions suggest rare scenarios worth targeted collection or synthetic augmentation.

# Compute embeddings
dataset.compute_embeddings(model, embeddings_field="embeddings")
# Generate 2D visualization using UMAP
results = fob.compute_visualization(
    dataset, 
    embeddings="embeddings",
    brain_key="img_viz"
)
# Launch interactive exploration
session = fo.launch_app(dataset)

Implement progressive annotation strategies

Rather than annotating entire datasets up front, adopt iterative expansion:

1. Initial selection: Curate 10% to 20% of the most unique/representative samples with coreset selection, mistakenness computation, or another algorithmic tool.
2. Auto labeling and training: Annotate quickly with foundation models and train your initial model from those labels.
3. Failure analysis: Identify prediction errors and edge case gaps.
4. Targeted expansion: Collect or annotate specific scenarios addressing weaknesses.
5. Iterate: Repeat cycle, focusing resources on high-impact improvements.

This approach mirrors active learning but with explicit curation intelligence guiding selection.

Automate quality gates

Replace subjective manual review with deterministic quality gates. Automated checks are the only way to catch systematic errors like schema violations or class imbalance that human reviewers inevitably miss at scale.

from fiftyone import ViewField as F
# Find bounding boxes that are impossibly small
tiny_boxes = dataset.filter_labels(
    "ground_truth",
    (F("bounding_box")[2] * F("bounding_box")[3])  0.8)

# Schema Validation: Find detections missing required attributes
incomplete_labels = dataset.filter_labels(
    "ground_truth",
    F("occluded") == None
)

Maintain annotation provenance

Track curation decisions and annotation metadata to support iterative improvement. This provenance enables sophisticated analysis of which curation strategies yield the best model improvements and supports continuous workflow optimization.

# Grab the "most unique" sample from a curated view of unique smaples
most_confusing_sample = unique_view.first()

# Add sample-level provenance
most_confusing_sample.tags.append("curated_for_review")

# Set metadata on the specific labels (detections)
if most_confusing_sample.detections:
    for det in most_confusing_sample.detections.detections:
        det["annotator"] = "expert_reviewer"
        det["review_status"] = "validated"
    most_confusing_sample.save()

A unified platform for curation-driven workflows

Voxel51’s flagship open source computer vision platform, FiftyOne, provides the necessary tools to curate, annotate, and evaluate AI models. It provides a unified interface for data selection, QA, and iteration.

Architecture advantages

Open-source foundations provide transparency into data processing while enabling customization for specific workflows. FiftyOne has millions of community users and an extensive integrations framework that lets you integrate FiftyOne with any workflow or external tool.

The design recognizes that curation, annotation, and evaluation are interconnected activities requiring shared context rather than isolated tools. This architectural philosophy enables the feedback loops that make curation-first workflows effective: evaluation insights immediately inform curation priorities, which drive targeted annotation, and which in turn feed back into refined models.

• Data-centric selection: Zero-shot coreset selection, uniqueness scoring, and embedding-based exploration enable intelligent prioritization before any annotation investment.
• Unified annotation: Create and modify 2D bounding boxes, 3D cuboids, and polylines directly within the platform where you already curate and evaluate. Annotate and QA 2D and 3D annotations in a single interface to maintain spatial context across modalities. (View a demo video.)
• ML-powered quality control: Mistakenness scoring, similarity search, and embedding visualization surface labeling errors systematically rather than through random sampling.
• Production-grade features: Dataset versioning captures state at each training iteration, annotation schemas enforce consistency, and programmatic quality gates prevent drift.

Getting started

Teams can implement curation-first workflows incrementally:

pip install fiftyone
# Load existing dataset
import fiftyone as fo
dataset = fo.Dataset.from_dir(
    dataset_dir="/path/to/data",
    dataset_type=fo.types.ImageDirectory
)
# Generate embeddings
model = foz.load_zoo_model("clip-vit-base32-torch")
dataset.compute_embeddings(model)
# Compute 2-D visualization
fob.compute_visualization(
    dataset,
    embeddings=embeddings,
    brain_key="clip_viz",
)
# Visualize and curate your data
session = fo.launch_app(dataset)

Future outlook: From reactive labeling to proactive intelligence

Three technical shifts are accelerating the move to curation-first workflows.

1. Foundation models as curators: Pre-trained vision-language models (VLMs) can now describe and filter images semantically without task-specific training. Instead of waiting for human review, teams can use multi-modal models to auto-tag complex sensor data (LiDAR/camera) and prioritize scenarios based on deployment needs.
2. Active learning meets intelligent curation: Standard active learning can waste budget by blindly flagging “low-confidence” predictions that are really just noisy or redundant frames. Next-generation pipelines now filter these requests through a uniqueness check. By prioritizing samples that are both confusing to the model and unique in the dataset, teams maximize the learning value of every labeled image.
3. Continuous curation in production: As models deploy to production, curation intelligence will extend to monitoring and maintenance. Embedding analysis of production data will detect distribution drift, trigger targeted data collection for new scenarios, and prioritize annotation of examples where models fail. This closes the loop from deployment back to development, enabling continuous model improvement grounded in real-world performance data.

Make your annotation investments count

Curation-first workflows coupled with smart labeling fundamentally transform how teams develop computer vision systems. Progressive annotation strategies focus on high-impact data help teams achieve better model performance with 60% to 80% less labeling effort.

For teams ready to make that shift, the path forward starts with understanding your data before you label it.

—

New Tech Forum provides a venue for technology leaders—including vendors and other outside contributors—to explore and discuss emerging enterprise technology in unprecedented depth and breadth. The selection is subjective, based on our pick of the technologies we believe to be important and of greatest interest to InfoWorld readers. InfoWorld does not accept marketing collateral for publication and reserves the right to edit all contributed content. Send all inquiries to doug_dineley@foundryco.com.

Developers can spend days using fuzzing tools to find security weaknesses in code. Alternatively, they can simply ask an LLM to do the job for them in seconds.

The catch: LLMs are evolving so rapidly that this convenience might come with hidden dangers.

The latest example is from researcher Hung Nguyen from AI red teaming company Calif, who, with simple prompts to Anthropic’s Claude Code, was able to uncover zero-day remote code exploits (RCEs) in the source code of two of the most popular developer text editors, Vim and GNU Emacs.

Nguyen started with Vim. “Somebody told me there is an RCE 0-day when you open a file. Find it,” he instructed Claude Code. 

Within two minutes, Claude Code had discovered the flaw: missing critical security checks (P_MLE and P_SECURE) in the tabpanel sidebar introduced in 2025, and a missing security check in the autocmd_add() function.

Claude Code then helpfully tried to find ways to exploit the vulnerability, eventually suggesting a tactic that bypassed the Vim sandbox by persuading a target to open a malicious file. It had gone from prompt to proof-of-concept (PoC) exploit in minutes.

“An attacker who can deliver a crafted file to a victim achieves arbitrary command execution with the privileges of the user running Vim,” Vim maintainers noted in their security advisory. “The attack requires only that the victim opens the file; no further interaction is needed.”

GNU Emacs ‘forever-day’

Surprised, Nguyen then jokingly suggested Claude Code find the same type of flaw in a second text editor, GNU Emacs.

Claude Code obliged, finding a zero-day vulnerability, dating back to 2018, in the way the program interacts with the Git version control system that would make it possible to execute malicious code simply by opening a file.

“Opening a file in GNU Emacs can trigger arbitrary code execution through version control (git), most requiring zero user interaction beyond the file open itself. The most severe finding requires no file-local variables at all — simply opening any file inside a directory containing a crafted .git/ folder executes attacker-controlled commands,” he wrote.

One fixed, one not

When notified, Vim’s maintainers quickly fixed their issue, identified as CVE-2026-34714 with a CVSS score of 9.2, in version 9.2.0272.

Unfortunately, addressing the GNU Emacs vulnerability, which is currently without a CVE identifier, isn’t as straightforward. Its maintainers believe it to be a problem with Git, and declined to address the issue; in his post, Nguyen suggests manual mitigations. The vulnerable versions are 30.2 (stable release) and 31.0.50 (development).

Vulnerable code

What does the discovery of these flaws tell us? Clearly, that large numbers of old codebases are potentially vulnerable to the power of AI tools such as Claude Code. Just because a weakness hasn’t been noticed for years doesn’t mean it will hide for long in the AI era.

That is, potentially, a big change, although hardly one that hasn’t already been flagged by Anthropic itself. In February, the company revealed that its Opus 4.6 model had been used to identify 500 high-severity security vulnerabilities.

“AI language models are already capable of identifying novel vulnerabilities, and may soon exceed the speed and scale of even expert human researchers,” it said at the time.

The platform is powerful enough that an enterprise version with the same capabilities, Claude Code Security, even negatively affected stock market sentiment towards several traditional cybersecurity companies when it was launched.

A second issue is that LLMs are now capable of spotting, iterating, and creating PoCs for vulnerabilities in ways developers still need to come to terms with. Meanwhile, the potential for malicious use is hard to ignore.

“How do we professional bug hunters make sense of this?” Nguyen asked. “This feels like the early 2000s. Back then a kid could hack anything, with SQL Injection. Now [they can] with Claude.”

This article originally appeared on CSOonline.

Meta researchers have developed a structured prompting technique that enables LLMs to verify code patches without executing them, achieving up to 93% accuracy in tests.

The method, dubbed semi-formal reasoning, could help reduce reliance on the resource-heavy sandbox environments currently required for automated code validation.

The development comes as organizations look to deploy agentic AI for repository-scale tasks like bug detection and patch validation. Traditional execution-based approaches often struggle to scale across large, heterogeneous codebases.

Instead of using free-form reasoning that can lead to hallucinations, the technique introduces structured logical certificates. These require models to explicitly state assumptions and trace execution paths before deriving a conclusion.

The researchers evaluated the approach across three key tasks, including patch equivalence verification, fault localization, and code question answering, and found that semi-formal reasoning improved accuracy across all of them.

“For patch equivalence, accuracy improves from 78% to 88% on curated examples and reaches 93% on real-world agent-generated patches, approaching the reliability needed for execution-free RL reward signals,” the researchers said in the paper.

For code question answering, semi-formal reasoning reaches 87% accuracy, marking a nine-percentage point improvement over standard agentic reasoning. In fault localization, it boosts Top 5 accuracy by five percentage points compared to standard approaches.

How it works

Semi-formal reasoning occupies a middle ground between unstructured chat and rigid formal verification. While standard reasoning allows models to make claims without justification, this approach uses a predefined template that mandates a step-by-step process.

“Rather than training specialized models or formalizing semantics, we prompt agents with structured reasoning templates that require explicit evidence for each claim,” the researchers said.

They added that the “templates act as certificates: the agent must state premises, trace relevant code paths, and provide formal conclusions. The structured format naturally encourages interprocedural reasoning, as tracing program paths requires the agent to follow function calls rather than guess their behavior.”

In practice, this forces the model to behave like a developer stepping through code line by line.

Researchers said that in one case involving the Django framework, the structured approach revealed that a module-level function shadowed Python’s built-in format() function. While standard reasoning missed this nuance, the semi-formal analysis correctly identified that the code would fail.

Implications for enterprises

Analysts said semi-formal reasoning signals a shift from assistive AI to more accountable AI in software engineering, a distinction that could reshape how enterprises approach code review.

“Tools like GitHub Copilot have conditioned developers to interact with AI as a fast, fluent suggestion engine,” said Sanchit Vir Gogia, chief analyst at Greyhound Research. “You ask, it generates, you accept or tweak. The system optimizes for speed and plausibility. What it does not optimize for is proof.”

Semi-formal reasoning changes that dynamic. Instead of rewarding models for sounding correct, it requires them to demonstrate correctness by tracing logic and grounding conclusions. For developers, this shifts the focus from reviewing outputs to evaluating the reasoning behind them.

“The deeper implication is that code review itself starts to evolve,” Gogia said. “Historically, code review has been a human bottleneck tied to knowledge transfer and design validation as much as bug detection. In practice, it often fails to catch critical issues while slowing down integration. What we are seeing now is the early shape of a machine-led verification layer where the system traces logic and the human validates the outcome.”

The shift, however, is not without tradeoffs. Structured reasoning introduces additional compute and workflow overhead, raising questions about how it should be deployed in real-world development environments.  

“More steps, more tokens, more latency,” Gogia said. “In controlled experiments, this can be justified by higher accuracy. In real developer environments, this translates into slower builds, longer feedback cycles, and increased infrastructure spend. If this is applied indiscriminately, developers will bypass it. Not because they disagree with it, but because it gets in the way.”

There is also a technical risk. The researchers noted that while the structured format reduces guessing, it can also produce “confident but wrong” answers. In these cases, the AI constructs an elaborate but incomplete reasoning chain, packaging an incorrect conclusion in a convincing, highly structured format that may be difficult for a human to quickly debunk.

Everyone is worried about junior developers. What are all these fresh-faced computer science graduates going to do now that AI is writing all the code?  

It is a legitimate concern. 

It wasn’t that long ago that the best advice I could give an early-career person interested in software development was to go to a boot camp. Sure, they could go to college and get a four-year computer science degree, but that would be expensive, take a long time, and teach them a lot of theoretical but impractical things about computers. And they wouldn’t even be doing science. 

But a six-month boot camp? There they’d learn what they really need to know—what software development companies are really looking for. They’d learn practical coding techniques, proper bug management, design specifications, JavaScript and TypeScript, source control management, and continuous integration.  

When I was a hiring manager, it didn’t take long for me to realize that a boot camp graduate was often much more ready to hit the ground running as a junior developer than a computer science graduate. 

But of course, all that fell apart overnight. Suddenly, for a low monthly payment, I could have a tireless, diligent, eager, and highly skilled junior developer who can type a thousand words a minute and reason at the speed of light. The economics of that are simply too compelling. 

Juniors begat seniors

And so what is a budding software developer to do? Or more importantly, what is a software development company to do when they realize that all those senior developers who are using Cursor are actually going to retire one day?  

Up until about 10 minutes ago, those companies would hire these intrepid young whippersnappers and put them to work fixing bugs, writing the boring code that builds systems, and slowly but surely teaching them how systems work by having them learn by doing. One became a senior developer through the experience of writing code, seeing it run, and learning what works and what doesn’t. Eventually, wisdom would set in, and they’d become sage, seasoned developers ready to mentor the next generation of developers.  

Well, we are now skipping that part where you actually become wise. But wisdom is actually the critical thing in this grand process. The judgment to know what is good, what is effective, and what is needed is the very commodity that makes agentic coding work. The AI model writes the code, and we seasoned veterans determine if it is right or not. 

We seasoned veterans know if the code is right or not because we’ve written tons and tons of code. But humans aren’t writing tons and tons of code anymore. And here is where I’m going to say something that I think many of you will really not like: Code doesn’t matter anymore. 

What I mean is, code is a commodity now. Code that used to take months to produce can now be produced in minutes. Yes, literally minutes. And the coding agents today are the worst they will ever be. They are only getting better, and they will only produce cleaner and cleaner code as time marches on. At some point—and that point may already be here for many of you—we are just going to stop looking at code. 

What matters is whether or not the application, you know, actually works. And if you want Claude Code or Codex to write a working application for you, you need to be able to communicate with it effectively to get it to do what you want. And strangely, the way to communicate with it is to write clearly. 

Heads up, English majors

A couple of weeks ago, I wrote that Markdown is the new programming language, and that what makes for “good code” in Markdown is the ability to write clear and concise instructions. Who would have thought that the English department would suddenly be the key to developing good software? 

Right now, the agentic coding process goes something like: 

1. Describe the problem to Claude Code.
2. Monitor the code Claude writes to make sure it is good code.
3. Test the application to make sure it works correctly.
4. Refine and improve by iterating this process. 

Step 2? It’s already becoming unnecessary. These AI agents are already writing good code, and the code they write gets better and better every day. And it is trivial to tell them to improve the code that they have already written. Iterating to improve code quality takes mere minutes. Writing the code has literally become the easiest part of developing software. 

So my advice to the kids these days: Learn to write clearly and precisely. Learn how to understand systems and describe them and their use cases. Make sure you can succinctly describe what you need software to do. English majors take note. Hiring managers? You too.

WebAssembly, or Wasm, provides a standard way to deliver compact, binary-format applications that can run in the browser. Wasm is also designed to run at or near machine-native speeds. Developers can write code in one of the various languages that compile to Wasm as a target (e.g., Rust), and deliver that program anywhere Wasm runs.

But Wasm by itself isn’t enough. An application, especially one running in a browser, needs standardized and controllable ways to talk to the rest of the system. The WebAssembly specification doesn’t speak to any of that by design. It only describes the WebAssembly instruction set; not how programs using those instructions deal with the rest of the system.

That’s what the WASI standard provides—abstractions for using the host system, such as how to perform network and storage I/O, and using host resources like clocks or sources of entropy for PRNGs.

Until now, CPython has supported WASI, but not in a formally defined way. Nothing described how CPython would support versions of WASI (the spec), or the WASI SDK (an implementation of the spec). With PEP 816, the CPython team has formally defined how to support both the spec and the SDK going forward.

Ultimately, the new definition will make it easier to deliver Python apps in the browser or anywhere else Wasm runs. There are just a few things developers need to know to ensure they’re using Wasm correctly with Python under the new rules.

How Python has historically used Wasm

Most languages, such as Rust, compile to Wasm as a binary target. Because Python is interpreted—at least, the default CPython implementation works that way—it doesn’t compile to Wasm directly. Instead, the interpreter itself is compiled to Wasm, and Python programs are run on that Wasm version of the interpreter.

There are drawbacks to this approach. For one, it means you need a full copy of the interpreter and the standard library to run any Python program. There is as yet no mechanism to compile a Python program for Wasm that would either include a copy of the interpreter or make it self-contained.

Another big drawback: Any modules not written in pure Python can’t run in Wasm unless a Wasm-specific version of that module is compiled ahead of time. Unless you have a specially compiled version of, say, NumPy, you can’t use that module in Wasm.

Some of these issues are limitations of Python as a language. Its inherent dynamism makes it difficult to deploy a standalone program. Rust, by contrast, can compile to a single binary artifact for any supported target.

But some of these limits can also be attributed to the Wasm environment. For instance, many methods in the standard library aren’t available in Wasm enviroments because the WASI SDK doesn’t expose the needed interfaces for those methods. The more Python and other languages demand such things, the more likely they are to show up in the Wasm environment.

This is where it is useful for Python to be explicit about which versions it’ll use for both Wasm and its software development kit (or SDK) going forward. Each version of Python can then provide better guarantees about the Wasm features it supports.

Wasm support in Python: WASI and the WASI SDK

Wasm support involves two things: WASI and the WASI SDK. The difference between the two is a little like the difference between the Python language in the abstract and the CPython runtime. The former (WASI) is the spec for how Wasm programs interact with the host system, which can be implemented any number of ways. The latter (the WASI SDK) is the official implementation of that spec.

The WASI SDK is a modified version of the Clang compiler, which uses a library called wasi-libc. This gives programs written in C (and C API-compatible languages) access to WASI’s APIs for the host (storage, networking, timers, etc).

In theory, we should just be able to compile a given CPython release with the most recent WASI SDK at the time. But things aren’t that simple. For one, the SDK’s biggest component, wasi-libc, doesn’t guarantee it’ll be forward- or backward-compatible. Also, some versions of the SDK may cause buggy behavior with some versions of CPython. As developers, we want to know that this version of CPython works with this version of the SDK—or at least be able to document which bugs appear with any given combination of the two.

How future releases of CPython will use WASI

CPython has been available on Wasm since version 3.11, with Tier 2 and Tier 3 support. The more official wasip1 is the better-supported target, while the older emscripten standard is the less-supported version. But Tier 2 support has been confined to the WASI “Preview 1” set of system calls. And for the reasons already stated, the WASI SDK CPython uses is not necessarily the most recent version, either: it’s SDK version 21 for Python 3.11 and 3.12, and SDK version 24 for 3.13 and 3.14.

All of this will change with future releases of CPython, with a couple of hard rules in place for using WASI and its SDK:

1. Any version of WASI or the WASI SDK supported by a given CPython version by its beta 1 release will be the version supported for the lifetime of that CPython release. For instance, if CPython 3.15 uses version 0.3 of the WASI spec and version 33 of the SDK (these are arbitrary numbers), then that version of WASI and the SDK will be supported for that version of CPython until it is formally sunsetted.
2. Any changes to the version of the WASI spec or SDK used for a particular release requires approval from Python’s steering council. But this shouldn’t happen outside of some extraordinary set of circumstances—for instance, if a bug surfaced that made a given version of the SDK unusable with a given CPython release.

The benefits of WASI version guarantees for CPython

Going forward, developers can look forward to significant improvements to how Python will work with WASI:

1. It won’t only be easier for CPython developers to know which versions of WASI and the SDK to target. It will also be easier for the rest of the WASI ecosystem to determine which Python versions are compatible with various WASI and SDK editions.
2. Developers maintaining Python libraries with extension modules will have a better idea of how to compile those modules to Wasm for each Python point release. They will then be able to take advantage of newer WASI features sooner, knowing that a specific CPython will support them.
3. Developers can add WASI support to their projects for a given version of CPython sooner in each release cycle for the interpreter, as the WASI and SDK versions should be locked down by the first beta release.

IT leaders are setting their operations strategies for 2026 with an eye toward agility, flexibility, and tangible business results. 

Download the January 2026 issue of the Enterprise Spotlight from the editors of CIO, Computerworld, CSO, InfoWorld, and Network World and learn about the trends and technologies that will drive the IT agenda in the year ahead.

An Anthropic employee accidentally exposed the entire proprietary source code for its AI programming tool, Claude Code, by including a source map file in a version of the tool posted on Anthropic’s open npm registry account, a risky mistake, says an AI expert.

“A compromised source map is a security risk,” said US-based cybersecurity and AI expert Joseph Steinberg. “A hacker can use a source map to reconstruct the original source code and [see] how it works. Any secrets within that code – if someone coded in an API key, for example – is at risk, as is all of the logic. And any vulnerabilities found in the logic could become clear to the hacker who can then exploit the vulnerabilities.”

However, Anthropic spokesperson told CSO, “no sensitive customer data or credentials were involved or exposed. This was a release packaging issue caused by human error, not a security breach. We’re rolling out measures to prevent this from happening again.”

But it wasn’t the first time this had happened; according to Fortune and other news sources, the same thing happened last month.

Don’t expose .map files

Map files shouldn’t be left in the final version of code published on open source registries, where anyone can download a package; they can be sources of useful information for hackers.

According to developer Kuber Mehta, who published a blog on the latest incident, when someone publishes a JavaScript/TypeScript package to npm, the build toolchain often generates source map files (.map files). These files are a bridge between the minified/bundled production code and the original source; they exist so that when something crashes in production, the stack trace can point to the actual line of code in the original file, not to some unintelligible reference.

What’s available in these files? “Every file. Every comment. Every internal constant. Every system prompt. All of it, sitting right there in a JSON file that npm happily serves to anyone who runs npm pack or even just browses the package contents,” said Mehta.

“The mistake is almost always the same: someone forgets to add *.map to their .npmignore or doesn’t configure their bundler to skip source map generation for production builds,” Mehta said. “With Bun’s bundler (which Claude Code uses), source maps are generated by default unless you explicitly turn them off.”

Think of a source map as a file that shows what parts of minified computer code, which is not easily understandable to humans, are doing, shown in the human-readable source code, said Steinberg. For example, he said, it may indicate that the code in a specific portion of the executable code is performing the instructions that appear in some specific snippet of source code.

A source map can help with debugging, he added. Without it, he said, many errors would be identified as coming from a larger portion of code, rather than showing exactly where the errors occur.

The world learned of this incident when security researcher Chaofan Shou posted this message early Tuesday on X: “Claude code source code has been leaked via a map file in their npm registry!”, along with a link to the file.

A common error

Leaving source map files in a package “is an incredibly common mistake developers make quite often,” said secure coding trainer Tanya Janca. “In this specific situation, it is more serious than it would be somewhere else, mostly because of the incredibly high value of the intellectual property involved, and because now malicious actors can analyze the source code directly for vulnerabilities instead of having to reverse engineer it, which adds time, cost, and complexity.”

Ideally, Janca said, developers should harden their build environment, so they don’t ship debug information/features with production. She offered these tips to developers:

• disable source maps in the build/bundler tool;
• add the .maps file to the .npmignore / package.json files field to explicitly exclude it, even if it was generated during the build by accident;
• exclude the .maps files from the list of published artifacts in the continuous integration/continuous deployment environment;
• carefully separate debug builds from production builds if there are differences; even the comments could be incredibly sensitive.

A critical layer

Any exposure of source code or system-level logic is significant, because it shows how controls are implemented, commented Dan Schiappa, president of technology and services at Arctic Wolf. With this information exposed, the number of people who now understand how the model enforces behavior, manages access, and handles edge cases increases, he said.

“In AI systems, that layer is especially critical,” he added. “The orchestration, prompts, and workflows effectively define how the system operates. If those are exposed, it can make it easier to identify weaknesses or manipulate outcomes. Knowing that attackers are still discovering the most optimal ways to leverage AI means that in any instance where a tool could be compromised, there are likely cybercriminals waiting in the wings.”

In a quiet corner of GitHub better known for weekend experiments than paradigm shifts, Drona Reddy, a data analyst at Amazon US, has published a single markdown file that promises to cut Claude’s output token usage by more than half, not by changing code, but by reshaping the model’s behavior.

The file, called Claude.md and available under an MIT license, outlines a set of structured instructions that claim to reduce Claude’s output verbosity by about 63% without any code modifications.

These instructions impose strict behavioral constraints on the model, including limits on output length, emphasis on token efficiency and accuracy, controls on speculation, rules for typography, and a zero‑tolerance policy on sycophantic responses. They also simplify code generation and define clear override policies, effectively training the model to respond more concisely and deliberately.

Reducing output tokens

The rationale is straightforward: eliminate what Reddy describes as Claude’s “frivolous” habits, stripping out everything that isn’t strictly necessary. That means no automatic pleasantries like “Sure!” or “Great question!”, no boilerplate sign-offs such as “I hope this helps,” no restating the prompt, and no unsolicited suggestions or over-engineered abstractions.

It also curbs stylistic quirks like “em” dashes, smart quotes, and other Unicode characters that can break parsers, while preventing the model from reflexively agreeing with flawed assumptions.

At scale, that kind of austerity, according to Reddy, could translate into meaningful savings, turning small stylistic trims into outsized efficiency gains.

The data analyst also outlined three distinct use cases where the markdown file could be most effective. First, high-volume automation pipelines, such as resume bots, agent loops, and code generation, where verbosity compounds across repeated calls.
Second, repeated structured tasks, where Claude’s default expansiveness can add up over hundreds of interactions. Third, team environments that require consistent, parseable output formats across sessions, where tighter control over responses improves reliability and downstream usability.

In his own simulations on Claude Sonnet, Reddy said the file could save close to 9,600 tokens a day at 100 prompts, translating to roughly $0.86 in monthly savings. At 1,000 prompts a day, the savings rise to about 96,000 tokens, or $8.64 a month, while across three projects combined, he estimates reductions of nearly 288,000 tokens, equivalent to around $25.92 monthly.

However, the data analyst also warned that the file might be really ineffective, even counterproductive, in certain use cases, such as single one-off queries, fixing deep failures, or exploratory work where feedback is required, as the file itself consumes input tokens on every message.

“The CLAUDE.md file itself consumes input tokens on every message. The savings come from reduced output tokens. The net is only positive when output volume is high enough to offset the persistent input cost. At low usage it costs more than it saves,” Reddy wrote in the repository’s documentation.

Modest enterprise gains

Analysts do see enterprises and their CIOs benefitting from the markdown file, at least to a certain degree, especially as they struggle to balance spiraling inference bills and moving agentic or other AI pilots into production.

“A 63% token reduction can meaningfully lower inference costs and latency for enterprises running high-volume Claude workloads,” said Charlie Dai, principal analyst at Forrester.

The gains, however, may be more operational than transformative.

“For CIOs, this method offers some operational benefits as it improves output consistency, improves latency, and enforces basic token discipline, which can help in scaling automation,” said Pareekh Jain, principal analyst at Pareekh Consulting.

However, Jain pointed out that though this is a “useful tactical optimization”, it does not fundamentally change enterprise AI economics.

“In enterprise settings, the tactic is likely to translate into more modest savings because output tokens are only a portion of total usage as input context, retrieval, and agent orchestration typically dominate costs,” Jain said. “As a result, most enterprises would likely see single-digit savings rather than the headline number,” he added. The markdown file is designed to be model-agnostic and should work across large language models that can follow structured instructions, though Reddy noted he has not tested its effectiveness on local models such as those running on llama.cpp or Mistral.

The release of the Flexera 2026 State of the Cloud Report provides vital insights into how enterprises are navigating the constantly changing landscape of cloud computing and increasingly AI-driven workloads. The findings show that the enterprise cloud discussion has changed significantly in recent years.

What started as a focus on cost-cutting and simple lift-and-shift migration has evolved into a complex balancing act involving governance, innovation, and measurable business value. This shift is not only logical but also expected, especially considering the rise of AI and the shared experience of enterprises that have faced unexpected cloud costs and unchecked spending for years.

Let’s explore what the report shows, why these findings are expected given industry trends, and what enterprises can do to extract real value from their ongoing—indeed, increasing—cloud investments.

Moving beyond cost-cutting

The main finding from the 2026 State of the Cloud Report is that organizations are clearly moving past an era where “cost-cutting at all costs” was the primary cloud strategy. We see a notable 12% increase in organizations reporting value delivered to business units year over year, even as the emphasis on simple efficiency and savings has decreased by 6%.

This marks a turning point in cloud adoption maturity. For years, cloud evangelists and technology leaders highlighted the cost advantages of moving from on-premises systems to the public cloud, and rightly so. However, the initial enthusiasm faded for many CIOs and CFOs as complexities increased, such as cloud sprawl, unpredictable billing, or wasted resources. According to the report, about 29% of cloud spending is still wasted, a sobering figure for any CFO’s office, but this also marks an improvement compared to years of uncontrolled growth. Now, the focus is on making productive investments and how cloud services can significantly impact business results.

Wiser organizations are trying to understand the cost of each service, align spending with results, and develop the discipline to measure value at every level, from individual projects to large-scale digital transformation efforts.

Oversight and governance models

As cloud complexity increases, so does the need for controls and clear leadership. The State of the Cloud Report shows a growth in the adoption of cloud centers of excellence (CCOEs) and finops teams, with 71% of organizations now using CCOEs. More organizations have dedicated finops teams to advise, manage, and optimize cloud spend. Teams are also more active in overseeing SaaS and cloud software usage.

What unites these efforts is a shift toward centralized accountability. Enterprises have learned the hard way that finops, CCOEs, and asset management cannot function in isolation. Cloud cost management collapses when roles and responsibilities are scattered. The most developed organizations coordinate efforts across teams to maintain shared views on spending, usage, and business priorities.

Some enterprises may worry about adding bureaucracy, but establishing transparency and focus will balance innovation with risk management and connect cloud spending directly to organizational goals.

Opportunities and risks of AI

Perhaps the most important change in this year’s report is the rise in AI-driven workloads, especially the adoption of generative AI. The data is clear: 45% of organizations now use genAI extensively, up from 36% a year earlier. Companies see significant opportunities for innovation and gaining a competitive edge through AI. At the same time, they realize that AI workloads are flexible, can incur unpredictable costs, and are often billed under new consumption-based pricing models that can quickly escalate without strong governance.

While there’s excitement about AI’s potential, there’s also a rapidly growing consensus that strong governance and cost controls must be put in place before spending spirals far beyond projections. We’re starting to see organizations appoint dedicated AI governance leaders who can ensure that innovation is safe, scalable, and firmly tied to business value and operational accountability (unlike the early, chaotic days of cloud).

SaaS spending, too, has exploded, with the most common monthly range among respondents jumping to $200,000 to $500,000. This is a sharp increase from the $50,000 to $100,000-tier that dominated last year, driven largely by the proliferation of AI-powered features and complex, usage-based pricing models. The conclusion: Cloud bills won’t shrink anytime soon, but organizations are determined not to repeat past mistakes. The focus now is on ensuring every dollar has a tangible return.

Three ways to boost cloud value

If you’ve tracked cloud growth as long as I have, these shifts are both natural and expected. Ten years ago, cost savings were the main focus of every cloud business case. As companies moved beyond the initial adoption phase, issues like sprawl, waste, and sticker shock emerged. The rise of AI and its endless demand for compute and data have increased both the opportunities and risks. Enterprises are seeking a balanced approach that carefully considers innovation, costs, and business value.

Cloud cost management is evolving into a smarter, value-driven discipline that’s keeping pace with business priorities and fast-changing markets. Here are three recommendations for enterprises looking to increase value in their cloud services:

1. Double down on centralized governance. Build or strengthen your CCOE and finops capabilities, but don’t let these teams operate independently. Accountability and transparency must be shared and tied directly to business outcomes.
2. Bring AI governance into the fold early. Establish clear ownership, set measurable goals, and don’t hesitate to manage experimental projects with the same rigor as production workloads. AI innovation is too expensive and too important to leave to chance.
3. Treat finops as an ongoing discipline, not a one-time project. Invest in the people, processes, and tools that support continuous tracking, optimization, and reporting. Incorporate cost and value discussions into every cloud project’s life cycle.

The key message from this year’s report is simple: Enterprises are demanding clear value from their cloud and AI initiatives. That’s expected, and it’s exactly where mature organizations should be headed.

Front-end engineers usually think performance problems live in the browser. When a page feels slow, we inspect bundle size and rendering. When something breaks, we open the network tab. If users complain, we optimize components or tweak state management. For a long time, I approached production issues the same way, assuming the root cause had to exist somewhere inside the UI. Over time, however, I started noticing a pattern: many confusing ‘front-end’ problems were not actually caused by front-end code. 

A login flow would occasionally fail and then work on refresh. An API would be slow only the first time. A deployment fix would be live for me, but not for a user. Sometimes, the interface displayed outdated data immediately after release. These issues were not caused by typical JavaScript errors. They were influenced by infrastructure behavior, particularly in environments running on AWS. 

Front-end engineers don’t need to manage servers to be affected by them. Modern web applications are no longer a single application talking to a single server. They sit on top of distributed cloud systems, and those systems influence how a UI behaves. Understanding a few core AWS concepts does not turn a front-end developer into a cloud engineer, but it does make debugging faster and UI design decisions more realistic. 

The hidden gap between front end and the cloud 

Front-end and back-end teams usually interact through a simple contract: an endpoint. The front end receives a URL and consumes data from it. From the UI’s perspective, it is just a request returning JSON. Behind that URL, however, is often a chain of services including gateways, caching layers, routing systems and load balancing. 

Because these layers are invisible, front-end engineers may make assumptions that don’t always match how distributed systems behave. When an API responds slowly, we suspect inefficient code. When requests fail intermittently, we assume unstable networking. When behavior changes between users, we think state handling is incorrect. In practice, many of these behaviors are predictable consequences of the infrastructure itself. 

The result is that UI code frequently compensates for system behavior without understanding it. Developers add unnecessary retries, misleading error messages or extra loading states. Once you recognize how the cloud shapes responses, the behavior stops appearing random and starts appearing explainable. 

How cloud infrastructure changes front-end behavior 

CDN hosting and the “old UI after deployment” problem 

Most modern front ends are deployed as static files. The application is essentially a set of HTML, CSS and JavaScript bundles delivered to the browser. In AWS environments, these files are commonly served through a content delivery network backed by object storage. This improves performance because users receive files from a location geographically close to them rather than from a single centralized server. 

However, that performance improvement comes with caching. After a deployment, some users may still see the previous version of the interface. A hard refresh fixes it, and waiting a short time fixes it as well. This often feels like a failed deployment, but it is expected behavior. The network is doing what it was designed to do: reuse previously downloaded files to improve speed. In practice, this behavior often comes from a combination of CDN edge caching, browser caching and cache headers rather than a single caching layer. 

From a front-end perspective, this changes how releases should be handled. Deployment is no longer only about shipping new code; it is also about ensuring browsers and caching layers request updated files. Versioned filenames and cache-aware design become important front-end concerns. Understanding that the infrastructure intentionally preserves older assets makes these issues predictable instead of mysterious. 

Serverless APIs and the slow first request 

Another behavior front-end engineers commonly observe is that an API request can be unusually slow the first time and normal afterward. This can be confusing because the same endpoint suddenly becomes responsive without any code changes. 

This behavior occurs because the API runs on serverless compute. Instead of a constantly running server, the platform initializes an execution environment only when a request arrives. The initial request includes the startup time required to initialize that environment. Once active, subsequent requests respond quickly. 

For UI design, this distinction matters. A loading state designed around consistent response times may incorrectly display an error or timeout during a normal cold start. Users interpret this as a broken feature even though the system is functioning correctly. Recognizing that occasional long responses are architectural rather than faulty allows front-end developers to design more forgiving loading states and avoid unnecessary failures. Cold starts are infrequent under steady traffic but noticeable in low-traffic or sporadic workloads. 

Understanding this also changes debugging. Not every delay is caused by network speed or inefficient queries. Sometimes the system is simply initializing itself in response to real usage patterns. 

Distributed systems and intermittent failures 

One of the most difficult production issues to investigate is a problem that cannot be reproduced locally. An interface may work consistently for developers but fail for certain users. Requests occasionally return server errors and then succeed moments later. 

Cloud environments distribute traffic across multiple machines and sometimes multiple regions. During deployments or scaling events, some users may temporarily reach instances that are being replaced, warming up or failing health checks. The infrastructure is designed for availability, but brief inconsistencies are normal in distributed systems and eventual consistency models. 

This reality affects front-end reliability. Interfaces benefit from not assuming every request will succeed immediately. Instead, they should recover gracefully, allow safe retries and present clear feedback to the user. When the UI anticipates occasional failures, the application feels significantly more stable even when the back-end behavior has not changed. 

Recognizing these failures as systemic rather than accidental helps teams avoid spending time debugging code that is functioning as intended. 

Why this matters for front-end engineers 

Understanding cloud behavior changes how front-end engineers approach everyday work. Instead of assuming uniform response times and perfectly consistent data, developers begin designing for real conditions: cached responses, variable latency and temporary unavailability. 

This shift improves both debugging and design. Problems are diagnosed more quickly because the source is clearer, and user interfaces become more resilient. Loading states feel more natural, errors are more accurate and deployments cause fewer surprises. 

Front-end engineers do not need to configure infrastructure or manage environments. However, modern interfaces are the visible layer of a distributed system. Learning a small amount about how cloud platforms behave helps developers align UI behavior with system reality. 

Knowing a few AWS fundamentals does not make someone an operations specialist. It makes them a front-end engineer who understands the environment their application runs in, and that understanding often has a greater impact on user experience than additional front-end optimizations. 

Disclaimer: The views expressed in this article are my own and do not represent those of my employer. 

This article is published as part of the Foundry Expert Contributor Network.
Want to join?

Microsoft has given Azure many hats: a serverless platform for distributed applications, a host for security and identity services, a place for big data, and an alternative to running your own data centers and infrastructure.

It’s this last one that’s often forgotten since much of the thinking about cloud platforms focuses on new tools and technologies instead of the old faithful applications that have been lifted and shifted to the cloud. The lift-and-shift process has become increasingly important as a tool to help get rid of servers and also make some headway against the perennial problem of technical debt.

Solving technical debt

Building new applications is easy, but it’s hard to keep them up-to-date. The budget for maintenance never quite covers the necessary work, so applications and servers keep running until it’s impossible to keep them going. Historically that’s meant aged mainframes running for more than 50 years, or a nuclear power plant being controlled from an emulation running in another emulation on top of a stack of virtual machines.

That’s the extreme end of technical debt—software archaeology rather than software engineering. Day-to-day technical debt is perhaps better thought of as a drag on business, code that doesn’t quite fit the current state of a business process, requiring manual workarounds that slow things down but allow the applications to keep running. There’s associated compounding risk as applications, operating systems, storage, and networks drift away from security baselines, dropping out of support, unpatched because any updates could stop the business from operating.

Migrating to the cloud can help, but too often it’s a matter of simply replicating a physical infrastructure in virtual machines on an IaaS platform. Yes, hyperscale cloud PaaS features could solve problems, as well as using automated updates and upgrades to deploy the latest features and keep applications more secure.

Automating cloud migrations

Microsoft has offered several generations of tools to migrate applications to the cloud, mostly focused on finding the right-size Azure virtual machines and the right virtual network appliances and topography, as well as importing data into cloud storage. Microsoft has provided useful ways to move applications to the cloud, but the company hasn’t helped you modernize code or infrastructure. Enterprises are just replicating technical debt in Azure instead of on their own servers.

Even if you use these tools, a migration can take months of planning and testing, and the disconnect between IT teams focusing on infrastructure and development teams focusing on the code means there’s no shared view of the entire task.

How then can Microsoft break down the barriers between teams and use migration as an opportunity to help modernize software and reduce technical debt?

Azure Copilot for migration and modernization

The latest version of the Azure Copilot recently launched, building on earlier releases and the agent model implemented in the closely related GitHub Copilot. It includes a new migration agent as part of its library of tools, using grounded AI to guide you through a simultaneous process of migrating and upgrading applications. The intent is to speed up the process by using your current environment and the capabilities of the Azure platform to define and implement the IT strategies to deliver a modern cloud infrastructure.

With the Azure Copilot handling infrastructure, the modernization tools in GitHub Copilot’s agents can help work through the necessary steps to update the code, adding support for Azure capabilities and modern cloud-native architectures.

Key to both approaches is a process of fine-tuning and grounding that uses Azure’s well-defined APIs and the constraints of domain-specific, software-defined infrastructure through tools like Bicep, Terraform, and the older but still critical Azure Resource Manager. Defining and implementing an IT strategy is one of the things an AI agent should be good at. Working alongside infrastructure and application architects, the agent defines the current state of an application, the target, where the modernized version will run, and what tools it will need. It then uses spec-driven development methodology to treat that gap as a directed graph that will first define an infrastructure and then update the code.

Having a known state at both ends of the process keeps risk to a minimum, though of course you need to always keep humans in the loop. It’s not a process that can be fully automated; instead, it’s an approach that speeds up tasks and reduces the necessary effort. In one early test, a customer was able to reduce this by 70%.

Working with cooperating agents

Microsoft is using migration as an example of how agents can cooperate and help different disciplines communicate effectively. Reports generated by GitHub Copilot can be used by Azure Copilot to identify possible issues and bridge the gaps between software modernization and migration plans. The resulting insights help teams prioritize necessary work and improve the specifications and strategy that guide the work.

Using the migration agent from Azure Copilot is straightforward, as it builds on existing best practices and processes. However, don’t expect it to support all the possible migration scenarios from day one. The current preview release is designed to help move specific infrastructures to Azure by analyzing data from the existing Azure Migrate tools.

Two key scenarios are supported in this first version: moving VMware infrastructures, and working with existing environments that use Hyper-V and physical servers. In both cases, you’ll need to run the existing Azure Migrate tools to collect the data the agent will use to plan a migration. This will require installing Azure Migrate collectors or the free RVTools utility. Microsoft provides an Azure Migrate appliance that can be deployed inside either VMware or Hyper-V environments (or on bare-metal servers) to handle discovery, which helps gather and process this data.

The migration agent will run discovery for you or work with your own discovery data. Once you have data, you can use prompts to assess your infrastructure, check for servers that need upgrades, and build a plan for a lift-and-shift exercise. You can even get cost analysis and ROI reports. Other options help add modernization options, for example, moving data to Azure servers. Then you can start building the base infrastructure for a migration and start deploying Azure resources.

Agents connect ops and dev teams

A conversational approach to working with the agent through Azure Copilot can help financial and business team members understand the effectiveness of a migration, as they can get access to costs and timescales through familiar tools. System administrators will be able to quickly get the information they need, while development teams will be able to understand what code changes might be needed to support a new infrastructure. Having Azure Copilot as a hub for these conversations can help reduce risks and keep projects on track. The information needed for good decisions is now easily accessible.

At the same time, you can have the GitHub modernization agent from the GitHub CLI update the code you’re running on those servers, using the tool to guide updates to .NET and Java. The agent will analyze code and produce a modernization plan to guide development teams, or it can automate the process of updating and testing your code. The migration agent is designed to look for issues that might arise when migrating to the cloud, so it’s an important component of a suite of migration tools.

With these new AI-powered tools, you’re able to speed up the process of moving complete applications to the cloud, with an ROI assessment, a migration plan, and the necessary updates to what may be outdated code. With new infrastructure and code, you’re able to start dealing with long-term technical debt and adding new features that can improve business performance and offer new services both inside and outside your organization.

Since its initial release in 2011, Apache Kafka has cemented itself as the de facto platform for event streaming. Kafka enthusiast Tim Berglund often refers to it as the “universal data substrate.” This is made possible in large part by the Kafka ecosystem that enables connectivity between Kafka and external systems (Kafka Connect) and a Java stream processing library (Kafka Streams).

The latest release of Apache Kafka delivers the queue-like consumption semantics of point-to-point messaging. After many hours of development and testing in recent releases, this feature is generally available in Kafka 4.2.

Let’s start with a quick “compare-and-contrast” of event streaming and message queuing. Event streaming is for high-volume, real-time processing of an unbounded, continuous stream of data, and it allows for consumers to replay old events as needed. Consumer applications record the offset (the ordinal position in each topic partition) of the last event Kafka successfully processed. If a consumer terminates or restarts, it’s able to resume processing the assigned partition from the last committed offset. Example use cases include internet ad attribution, updating ride-share status, and monitoring for credit card fraud. This is the space where Kafka has thrived, with adoption by over 80% of all Fortune 100 companies.

Message queues are used for point-to-point communication, where a message is typically consumed once and removed from the queue. Unlike with event streaming, consuming applications are able to acknowledge each message. This messaging pattern decouples applications and services via guaranteed, one-time processing for tasks such as in-app notifications to mobile devices, generating payroll records, or calling an AI model. Popular platforms in this space include RabbitMQ, ActiveMQ, and IBM MQ.

These message queue use cases have been a “square peg in a round hole” for Apache Kafka. Why? For starters, scaling the “traditional” Kafka consumer group is constrained by the number of topic partitions. Most notably, Kafka consumers don’t have message-level acknowledgement semantics. These features enable consumers’ message queue systems to cooperatively operate on messages in a queue.

This is the major motivation behind KIP-932: Queues for Kafka. Let’s see how this Kafka implementation of message queuing could be an important tool in your event-driven architecture.

Scaling Kafka consumer applications

Traditionally, parallel processing of Kafka topic data is constrained by the number of partitions of the topic being consumed. The broker assigns consumption of each partition of the topic to a single member of a consumer group. Once the membership of the consumer group equals the partitions of the topic, any new consumers added to the group will be idle.

Confluent

KIP-932 adds a new type of group called a share group. Nothing changes about how the data is written to Kafka by producer applications or how data is stored in Kafka. Your event streaming use cases can operate on the same topics.

Share groups introduce a new cooperative consumption model, where consumers in a share group work in a similar fashion to consumers/subscribers in message queuing systems. On the broker, each topic-partition has a corresponding share partition which tracks the lifecycle of each message in relation to the share group. This allows the share consumers to be scaled beyond the number of topic partitions.

Confluent

This cooperative consumption from a topic partition also means we lose the partition-level processing order guarantees of the “traditional” Kafka consumer. That’s the trade-off for this scaling, but cooperative consumption also is intended for use cases where throughput and scaling take precedence over the order of processing.

Message-level acknowledgement

The APIs for KIP-932 should be familiar to developers who are already using Kafka. For starters, nothing changes about how events are produced to Kafka topics. On the consumer side, the KafkaShareConsumer interface is very similar to the existing KafkaConsumer. Consumer applications will poll for available messages and process each resulting ConsumerRecord instance.

The consumers now have the ability to acknowledge the delivery of each record on an individual basis. By default, every message is implicitly acknowledged as successfully processed. However, there are scenarios where the developer needs more fine-grained controls, particularly around error handling and long-running tasks.

By using the value of explicit for the consumer configuration’s share.acknowledgement.mode, the code takes on the responsibility of specifying how each message should be acknowledged. The available AcknowledgementType values are ACCEPT, RELEASE, REJECT, and RENEW. These values influence the state of each message in relation to the share group. Those states are AVAILABLE, ACQUIRED, ACKNOWLEDGED, and ARCHIVED.

Confluent

Only messages in an AVAILABLE state can be fetched by a consumer. When fetched, a message transitions to the ACQUIRED state and a delivery count for that message is incremented. This effectively “locks” this message from fetches by other members of the share group.

Once ACQUIRED, a message is expected to be processed in a finite amount of time. If this “lock” or “lease” expires, the message is either sent back to the ACQUIRED state or moved to an ARCHIVED state, based on the delivery count of the message. The state and delivery count of each message is tracked in the share partition. This provides for a built-in retry mechanism developers can use in the event of a condition where the message process could be reattempted, as the message could be acknowledged using the RELEASE type.

If message processing completes successfully, that message is acknowledged with the ACCEPT type. This transitions the message to the ACKNOWLEDGED state.

There are cases where processing takes a non-deterministic amount of time. Perhaps the consumer calls a third-party or partner API. Maybe it’s augmenting the message with the result of an LLM call. These aren’t “failures,” and the processing code may need more time to complete. In this case, acknowledge the message with the RENEW type to reset the lock.

Unifying messaging protocols and infrastructure

Many organizations have both event streaming and message queuing use cases. This often means operators are maintaining and supporting Apache Kafka and an older message queuing system. Developers integrate applications with different messaging libraries and protocols in the same application code base. All of this happens as the C-suite is asking why we’re paying for multiple messaging solutions.

Consolidating these messaging use cases onto Apache Kafka will make producing applications simpler to develop, deploy, upgrade, and maintain. It will also help consumer applications scale to meet the needs and SLAs of the messages being processed.

Unlike traditional message queue systems, events in these “queues” enjoy the durability and storage guarantees we’ve come to rely on in Apache Kafka. Developers of consumer applications determine if the events should be processed as event streams or queues.

Operators and SREs (site reliability engineers) tend to like simplicity. (That could be due to the correlation between simplicity and the number of production incidents.) Unifying these messaging platforms means fewer systems to configure, deploy and patch. And that also addresses the concerns of the C-suite — lowering the total cost of ownership for the overall application infrastructure.

What queues for Kafka means for teams

KIP-932 brings long-awaited point-to-point semantics to Apache Kafka. This implementation layers queue-like consumption and message-level acknowledgment onto the durability, scalability, and throughput that have made Kafka mission-critical infrastructure for businesses from startups to large enterprises.

For development teams, this means writing applications against a single messaging API rather than juggling multiple protocols. For operations teams, it means consolidating infrastructure and reducing complexity. And for organizations, it means lower total cost of ownership without sacrificing the specific semantics each use case requires.

KIP-932 is available in Apache Kafka 4.2 and Confluent Cloud, with support coming to Confluent Platform version 8.2. Developers can explore the implementation and start testing queue-based consumption patterns now. For more about KIP-932 and other event streaming topics, visit Confluent Developer for free learning resources curated by our team of experts.

—

New Tech Forum provides a venue for technology leaders—including vendors and other outside contributors—to explore and discuss emerging enterprise technology in unprecedented depth and breadth. The selection is subjective, based on our pick of the technologies we believe to be important and of greatest interest to InfoWorld readers. InfoWorld does not accept marketing collateral for publication and reserves the right to edit all contributed content. Send all inquiries to doug_dineley@foundryco.com.

Anthropic didn’t intend to introduce Mythos this way. Details of what it calls its most capable AI model yet surfaced through a data leak in its content management system (CMS), revealing a LLM with sharply improved reasoning and coding skills.

The data leak, which was the result of the company’s staffers inadvertently exposing material about the LLM, including a draft blog post about it, via a publicly accessible data repository, was first identified by independent security researchers last week.

Following disclosure of the issue, Anthropic restricted public access to the data store, only to later attribute the exposure to a configuration error in its CMS and confirm the existence of the model to Fortune, which was the first to report the leak.

Apple-focused leaker M1Astra also flagged the exposure, archiving a copy of a draft Anthropic blog post about Mythos on X before access was restricted.

In that draft, Anthropic itself struck a cautious tone, signaling concern about the model’s potential implications on cybersecurity.

“In preparing to release Claude Mythos, we want to act with extra caution and understand the risks it poses — even beyond what we learn in our own testing,” the company wrote, adding that it is particularly focused on assessing near-term cybersecurity risks.

The blog further stated that Anthropic wants to seed Mythos across enterprise security teams first and has already been testing the model’s cybersecurity prowess with a “small number of early access customers.”

The rationale seems straightforward: if today’s models can already identify and even help exploit software vulnerabilities, a more capable system like Mythos could significantly accelerate both discovery and misuse — raising the stakes for defenders and attackers alike.

Pareekh Jain, principal analyst at Pareekh Consulting, says Mythos could cut both ways for CISOs and enterprise security teams, compressing the gap between cyber offense and defense.

While at one end, models like Mythos could transform security by automating vulnerability discovery, continuous red-teaming, faster triage, and large-scale threat hunting areas, on the other hand, it could make cyberattacks easier by letting AI agents act autonomously with high skill, Jain said.

That risk for CISOs is not theoretical, Jain added, as earlier-generation models were quickly repurposed into tools for developing malware.

The risk is even higher with Mythos because of its capabilities like “recursive self-fixing,” Vladimir Belomestnov, senior technical specialist at HCLTech, wrote in a post on LinkedIn.

“The leaked files highlight a capability for the AI to autonomously identify and patch vulnerabilities in its own code. Even if this is currently limited to assisted exploitation, it suggests a narrowing gap between human and machine software engineering,” Belomestnov wrote.

However, Anthropic appears to be some distance from a full release of the model.

“Mythos is also a large, compute-intensive model. It’s very expensive for us to serve, and will be very expensive for our customers to use. We’re working to make the model much more efficient before any general release,” the copy of the draft blog post reads.

What is clear, however, is that the company is already planning a phased rollout targeting cybersecurity use cases.

“We’ll be slowly expanding access to Claude Mythos to more customers using the Claude API over the coming weeks. Since we’re particularly interested in cybersecurity uses, that’s where we aim to expand the EAP initially,” the company wrote in the draft blog post.

There is another copy of the blog post, which also names the model as Capybara. Anthropic hasn’t made it clear what the final name of the model will be.

The indecision over the model’s name, though, didn’t stop it from rattling markets last week. Shares of cybersecurity vendors, including CrowdStrike, Palo Alto Networks, Zscaler, and Fortinet, fell as investors assessed what more capable models within Claude Code Security could mean for the competitive landscape.

However, Avasant’s research director, Gaurav Dewan, was more optimistic about Mythos’ impact on vendors: “Powerful models will not replace cybersecurity platforms”.

Rather, Dewan sees vendors increasingly embedding frontier models from Anthropic and OpenAI and others into their stacks for vulnerability discovery, code and cloud posture management, and threat investigation and response automation.

“One can expect partnerships and controlled integrations, not disintermediation. Vendors that already own telemetry, workflows, and enforcement will benefit most,” Dewan added.

The article originally appeared in CSO.

Paraphrasing William Gibson, the future of AI is here, but it’s nowhere close to evenly distributed yet.

Last week in London, I had two conversations about enterprise AI that obliterated any semblance of a neat and tidy story of AI adoption. In the first meeting, the head of engineering at a large hedge fund told me about engineering teams with fleets of agents in full production, and (in his personal case) all code is written by LLMs. (Junior hires, interestingly, aren’t allowed to use LLMs for code assistance.) In another meeting, a data engineer at a large retail bank described the exact opposite: No agents and sparse use of LLMs. Maybe other parts of the bank are moving faster on AI, but his division clearly isn’t.

This isn’t about one company “getting” AI and the other not. Rather, it’s a reminder that even within the same company there are wildly divergent adoption curves for new technologies. AI is widening the gap between teams that can absorb it operationally and teams that can’t. That’s what the best recent data suggests, too. McKinsey found that 88% of respondents say their organizations are using AI in at least one business function, but only about one-third say their companies have begun scaling AI programs. As for agents, 23% report scaling an agentic AI system somewhere in the enterprise, while 39% are still just experimenting. And in any given function, no more than 10% say they’re scaling agents.

Broad usage, in other words, is not the same thing as deep institutional change. In short, there’s still time to figure out AI. You’re not behind.

Cue the engineering boom

I keep hearing that “finance is cautious” or “regulated industries are behind” or “everyone is building with agents.” None of that is quite true. Some financial firms are moving aggressively. Some aren’t. Some teams inside the same firm are doing both at once. Deloitte’s 2026 enterprise AI research makes the same point from another angle. Only 25% of respondents said they had moved 40% or more of their AI pilots into production. Just 34% say they’re using AI to deeply transform their businesses (a number I suspect is more aspirational than actual), while 37% are still using it at a surface level with little or no change to core processes. That sounds a lot less like a tidal wave and a lot more like a messy, uneven organizational test.

Same as it ever was, right?

And that, in turn, is why I think a lot of the “AI will wipe out software jobs” talk is wrong and misses the point. The interesting thing about AI coding tools isn’t that they make software cheaper to produce. It’s what companies do with that lower cost. Box CEO Aaron Levie recently invoked Jevons paradox to explain exactly this dynamic: When a capability becomes cheaper and easier to consume, demand for it often rises rather than falls. That’s not a law of nature, but it is a pretty good description of what technology has done for…ever. Cloud computing didn’t lead companies to need less compute. It made them build more things that consumed compute. AI-assisted coding may be doing something similar for software itself.

This is where the data on engineering jobs gets interesting. Lenny Rachitsky recently highlighted that engineering openings are at their highest levels in more than three years. The underlying TrueUp data shows 67,665 open engineering jobs as of March 2026, up 78.2% from the recent low. More importantly, this isn’t just concentrated at the very top end of the market. TrueUp’s breakdown shows 44.6% of posted engineering roles within tech companies are entry and mid-level, versus 38.3% at senior level and 13.8% at senior-plus. So no, the data doesn’t say AI is eliminating roles for junior developers; rather, it says companies still want a lot of engineers, even as AI tools spread throughout the profession.

There’s a cleaner way to understand what’s happening. AI isn’t killing the need for engineers. It’s changing what enterprises want from engineers.

Stack Overflow’s 2025 survey found that 84% of respondents are using or planning to use AI tools in development, and just over half of professional developers use them daily. McKinsey’s software development research found that the highest-performing AI-driven software organizations are seeing 16% to 30% improvements in productivity, customer experience, and time to market, along with 31% to 45% improvements in software quality. But McKinsey’s crucial point is that these gains don’t come from sprinkling copilots over an unchanged process. They come from reworking roles, workflows, and the full product development system. That’s a much harder organizational challenge than buying licenses for a coding assistant.

Software engineering is alive and well

Let’s go back to my conversations in London. The hedge fund leader may be an early glimpse of where parts of enterprise engineering are headed. Less time hand-authoring code, more time specifying, reviewing, steering, and orchestrating systems that increasingly generate code for you. But that does not mean the retail bank division is irrationally lagging. In a heavily regulated environment, code generation is not the hard part. Governance is. Deloitte reports that only 21% of surveyed companies currently have a mature governance model for autonomous agents (and those 21% are probably kidding themselves). At the same time, 73% cite data privacy and security as a top risk, and 46% cite governance capabilities and oversight. That’s not bureaucracy for its own sake. It’s a recognition that plugging non-deterministic systems into deterministic, compliance-heavy environments gets messy fast.

Still, caution isn’t free. Every quarter a team spends in pilot mode is a quarter in which more aggressive peers are building operational muscle. OpenAI’s enterprise usage data is useful here because it shows how uneven that muscle-building already is. Frontier workers, defined as the 95th percentile of adoption intensity, send six times more messages than the median worker. Frontier firms send twice as many messages per seat. OpenAI says the primary constraints are no longer model performance or tools, but rather organizational readiness and implementation.

This rings true to me. In my experience, the real divide is increasingly not between companies that have access to AI and those that don’t. It’s between teams that have learned how to integrate AI into repeatable work and teams that are still treating it as a promising but dangerous sideshow, as I’ve written.

This is also why I think the distinction of task versus job matters. Writing a chunk of boilerplate code is a task. Engineering is a job. Jobs bundle judgment, trade-offs, accountability, architecture, security, integration, testing, and the ugly reality of operating systems in the real world. AI can automate more tasks, but it hasn’t eliminated the need for jobs, especially in environments where bad software decisions carry real operational or regulatory consequences. In fact, McKinsey’s broader AI survey found that most organizations are still navigating the transition from experimentation to scaled deployment, and that high performers stand out precisely because they redesign workflows and treat AI as a catalyst for innovation and growth, not just efficiency. That is a very different thing from saying, “We gave everyone a chatbot and now we need fewer people.” (By the way, that would be a very naive statement.)

So no, AI isn’t plodding (or rocketing) toward one uniform enterprise future in which software engineers quietly fade away. Instead AI is splitting enterprises into fast-learning and slow-learning teams and is rewarding organizations that redesign work, govern risk, and turn lower software costs into more software, not less. The code may be getting cheaper, but the ability to decide what should be built, how it should fit together, and how to keep it from breaking the business keeps increasing in value.

That’s not the death of software engineering. It’s the repricing of it, and every company and every team is paying different prices.

Just as integration catalogs were must-haves at the peak of SaaS, Model Context Protocol (MCP) servers are now becoming all the rage for connecting AI agents and enterprise systems. 

In this paradigm, developers aren’t hand-coding API calls to external systems, nor are users clicking “click to integrate” and entering credentials into GUIs. Instead, agentic systems are looking up available MCP servers and making MCP tool calls autonomously. To guide this process, MCP registries are emerging as a way to catalog available MCP servers and guide AI agent workflows.

“MCP registries are increasingly becoming the integration catalog for agentic systems,” says Ebrahim Alareqi, principal machine learning engineer at Incorta, provider of an open data delivery platform. “They give developers and platform teams a centralized inventory of the tools, agents, and capabilities available to an organization.”

MCP registries shorten time to integration and act as a discovery point for AI agents. But whether you’re repurposing an off-the-shelf MCP registry or building your own, the job comes with daunting technical challenges, like figuring out semantics for tool discovery and adding in guardrails for safe autonomous usage.

“A good MCP registry is more than a directory of tools,” says Derek Ashmore, agentic AI enablement principal at Asperitas Consulting, a cloud computing consultancy. “It’s part of your control plane.” For Ashmore, an MCP registry needs strong identity and discovery, policy-aware metadata, life-cycle controls, security guardrails, and data to inform observability.

Below, we’ll dive deeper into what makes up a solid, functional MCP registry. We’ll explore the features and requirements of MCP registries, examine the emerging implementation advice for enterprises, and determine when you should use private versus public MCP registries.

What is an MCP registry?

An MCP registry acts as a single source of truth for MCP servers. It’s a catalog of approved, compliant MCP servers and MCP tools that are available within an organization that can be exposed to AI agents. By pointing AI agents to an MCP registry endpoint, an enterprise can equip AI workflows with actionable read-write access across engineering, business, and SaaS systems that are sanctioned and configured for company use.

To date, there are a handful of public MCP registries out there, ranging from open directories to curated registries and more enterprise-ready implementations. The most obvious example is the official MCP Registry, an open-source catalog of MCP servers with a live REST API for search and discovery. The official MCP registry aligns with the MCP registry specification, which provides a standardized method to build interoperable MCP registries.

Other public resources are more like static lists, such as directories from MCP.so, to Glama.ai, Mastra.ai, and OpenTools. Interestingly, one open-source tool, MCP-Get, provides a command-line option for interaction.

More and more digital services are beginning to embed MCP catalogs into their platforms, too. Docker and Microsoft, for instance, are building curated MCP catalogs focused on their own platform ecosystems. GitHub hosts a directory of MCP servers for easy installation, and has begun to add controls for internal registry configurations. The MACH Alliance, an industry consortium focused on composable commerce, is also promoting an MCP-compliant registry initiative.

Beyond these efforts, MCP registries are moving beyond public directories — enterprises are now constructing private, self-hosted registries for governed internal MCP use. While examples of private enterprise MCP registries are nascent, they are undoubtedly being aided by emerging features in infrastructure platforms.

Take the MCP Center, powered by the Azure API Center, which demonstrates how to build MCP registries in Azure. Lunar.dev’s Custom MCP Server Registry also allows admins to create their own scoped internal MCP registries.

The benefits of an MCP registry

“The biggest benefit of an MCP registry is discoverability,” says Justin O’Connor, founder at Infracodebase, an agentic platform for cloud infrastructure, which hosts a public MCP registry for connecting AI agents to cloud providers.

“MCP servers often end up scattered across teams and systems, so a registry gives you one clear place where people can find what exists,” says O’Connor. This allows AI agents to discover tools with less trial and error.

Others agree that improved discovery is a much-needed element for autonomous agents. “MCP is designed to ensure agents have enough context to generate the right response, and registries are a natural extension of that,” says Incorta’s Alareqi.

A well-constructed MCP registry brings uniformity that aids adoption, reuse, and governance, says O’Connor, because it can be treated as an official inventory of approved capabilities. As such, MCP registries serve a similar function to package registries for software, he adds. 

In this way, an MCP registry can act as a source to vet and update MCP servers before exposing them to agents. Although an MCP registry doesn’t replace core authentication requirements for each MCP server, it does aid provenance and supply chain security, says O’Connor.

Core elements of an enterprise-grade MCP registry

While they share similarities with traditional software integration catalogs, MCP registries have some unique elements. “If you are treating the AI or MCP registry as just another static catalog, you’re doing it wrong,” says Christian Posta, VP and global field CTO at Solo.io, a cloud-native infrastructure company.

Many elements make up a high-quality MCP registry beyond a static tool catalog. In general, they can be boiled down to rich tool metadata, features for developers, and enhanced security guardrails. The effectiveness of an MCP registry will also depend on underlying MCP and API security best practices.

Rich tool metadata

First, an MCP registry needs the bread-and-butter details required to function with MCP. “A solid MCP registry needs to support the basics required by the protocol,” says O’Connor of Infracodebase. This includes how to connect to a server, the transport type, server URL, and configurations required, like environment variables or secrets.

Next are details to aid tool discovery. An MCP registry must provide methods for AI agents to automatically discover the appropriate underlying MCP tools. 

According to Posta, making tools discoverable requires resources that enable semantic search, such as embeddings of the tool name, description, and input schema, along with clear summaries. Ideally, he adds, this experience layer supports progressive disclosure to optimize context windows.

“Agents need context,” adds Incorta’s Alareqi. “Metadata around capabilities, schemas, side effects, cost, latency, and failure modes, to name a few, is what allows an agent to choose the right tool.”

William Collins, director of tech evangelism at Itential, provider of an infrastructure orchestration platform, also sees semantic cues as necessary for discovery, along with other metadata. These should flag rich semantic metadata beyond endpoint descriptions, versioning with breaking-change signaling, and clear capability scoping, he says.

Developer controls

Although agents will use MCP registries programmatically, the registries still must be maintained by human developers (most likely by platform engineers). The MCP registry should therefore provide controls to add new servers, remove them, and set privileges.

To streamline this, a key aspect of an effective MCP registry is good developer experience, says Ido Halevi, director of product management at Silverfort, an identity security company. “That means clear documentation, examples of usage from other teams, and reliability signals such as active maintenance and adoption across agents,” Halevi says.

A strong registry also provides context beyond being a basic tool list. “Teams need to know whether an MCP server is maintained, how widely it’s used, and what kinds of risks or privileges it requires,” says Jessica Kerr, engineering manager of developer relations at Honeycomb, an observability platform provider. For instance, Kerr suggests adding lightweight moderation controls to flag dependable versus experimental MCP servers.

Security guardrails

Since the concept of MCP registries is so new, security standards and guidelines are still emerging. “It’s a bit like the wild west,” says Gil Feig, co-founder and CTO of Merge, provider of a unified API platform. 

Because of this, Feig emphasizes the need for strong security guardrails and privilege boundaries. “When evaluating an MCP registry, look for one that offers robust authentication, observability, and data governance with built-in rules, proactive alerts, and real-time logs,” he says.

The authorization context will especially matter to ensure that agents are using MCP tools permitted by the organization and have authorized access to sensitive material. As such, MCP registries will require information on the agent identity, its intent, and what user it’s acting on behalf of, says Posta. 

“Registries should favor servers that properly separate user sessions so data does not leak between users,” adds O’Connor, who notes that support for per-user authentication using modern OAuth patterns helps ensure access that is matched to privileges.

Similarly, Halevi underscores the need for enforcement beyond pure tool discovery. “Without enforcement, all you’re doing is cataloging risk,” he says. A registry should help control which agents can access which tools, and dynamically enforce permissions when a tool is invoked.

Native API handling underneath

Native API handling notwithstanding, there’s only so much a registry can do. Core authentication nuances will differ from MCP server to MCP server, and each will require the same security rigor as a standard API connection.

“At the server level, MCP servers must be built with robust security capabilities from the ground up,” says Alex Salazar, co-founder and CEO of Arcade.dev, the maker of an AI tool calling platform. An MCP registry doesn’t replace core MCP server security basics such as OAuth-based authentication, proper token and secrets handling, and observability.

“The issue here is many AI applications don’t have any native API handling in place,” adds Melissa Ruzzi, director of AI at AppOmni, a cybersecurity company. “So they look to the MCP registry as a way to control MCP authentication, which is not a good practice.” 

Others aren’t certain guardrails belong at the registry level to begin with. “Security guardrails and privilege boundaries are really the responsibility of the underlying agents and not the best function of a registry-as-exchange,” says Dan Fink, AVP software architect at Cognizant, an enterprise technology consulting firm.

To really enforce this, adds Fink, you’d require additional layers that would either be too heavy, like introducing entirely new agents as intermediaries, or just simple guardrail tags, which could easily be faked or obsoleted.

For this reason and others, some view the MCP registry itself as more of an abstraction layer, which only defines high-level capabilities that are then mapped to underlying scopes, roles, and APIs. 

“Registries should express guardrails so orchestration layers can enforce them,” says Itential’s Collins. “This way, the registry doesn’t become a bottleneck and single point of failure.” For Collins, guardrails to enforce at the registry include privilege boundaries, authentication requirements, and risk classifications.

“An enterprise MCP registry should be slightly abstracted, not one-to-one with every tool privilege,” says Asperitas Consulting’s Ashmore. A thin abstraction layer, as opposed to one that directly mirrors every underlying permission, also enables you to standardize permission names across tools, reuse role templates, and separate user types, he adds.

Life cycle and performance

As a tagalong to security guardrails, an MCP registry is an opportune location to introduce supply chain security features and monitoring.

“This includes vetting servers before they’re discoverable, implementing security scans and vulnerability checks, and controlling what can be published or discovered,” says Arcade’s Salazar. He says that registries should track performance metrics and errors, as well.

In addition to dynamic tool discovery and tooling governance, Marco Palladino, CTO and co-founder of Kong, provider of a cloud-native API platform, sees observability across the AI data path as necessary for an enterprise-grade MCP registry.

“Enterprises need centralized visibility into tool usage, health, and failures to support monitoring, optimization, cost management, and compliance,” says Palladino. “Without this, organizations face fragmented integrations and increased operational risk.” 

Beyond the above areas, experts foresee that other attributes will be necessary for MCP registries in an enterprise context: 

• Fingerprinting of the tools within a particular server
• A bridge between private and public registries
• Ranking or scoring based on previous performance, token cost, and other attributes
• Namespace verification to prevent naming conflicts
• Validation layers to catch errors
• Health monitoring to track server availability and performance 

Choosing a public or private MCP registry 

When implementing an MCP registry, organizations have two options: either use a public MCP registry or create a private self-hosted MCP registry. According to the experts, there are trade-offs between each approach.

“A public MCP registry has to be very well evaluated for possible security risks before use,” says AppOmni’s Ruzzi. Private registries are generally safer, she says, but the degree of risk depends on how they are implemented.

“The public registry ecosystem is still immature,” says Kevin Cochrane, CMO at Vultr, a cloud hosting provider. “We likely need a ‘Hugging Face for MCP’ — a trusted authority that can validate listings and set consistent standards.” Without that sort of layer, teams should be cautious about smaller third-party registries, he adds. 

Instead, a private MCP registry can help an enterprise govern its portfolio. “Put a private MCP registry at the heart of the AI runtime,” Cochrane says. “This should be core infrastructure owned by platform engineering, with governance over how MCP servers are built, tested, deployed, and monitored.”

Infracodebase’s O’Connor adds that such curated registries engender trust in specific tools. “Over time, registries also become a trust boundary, especially in public settings, because they shape what tools people are willing to bring into workflows,” he says.

For many, the starting point will likely be a combination of both. This could equate to forking a sample open-source MCP registry and extending it to your needs. 

“Another way is to take a published OpenAPI specification and generate a skeleton service implementation in a language of your choice,” says Andrei Denissov, associate director of software engineering at Cognizant AI Lab, the AI research arm for Cognizant.

Tips on building MCP registries

Experimentation with MCP registries is in its early days. However, developers on the front lines are already pulling out lessons learned and discovering patterns for both good and bad designs. 

One lesson is the sheer realization that you need registries, quicker than you think. “Working with teams deploying MCP at an enterprise scale, the pattern is consistent: Registries become necessary faster than organizations expect,” says Silverfort’s Halevi. 

Then, those implementing MCP registries quickly learn that a basic MCP catalog is only one part of the picture — enterprises need much more than just MCP tool discovery. They need per-agent authorization models, guaranteed human-linked attribution, deep observability into agent behavior, and inline enforcement,” says Halevi.

When operating many MCP servers at scale, other requirements beyond discovery begin to become just as important, adds Halevi, such as MCP server orchestration, managing keys, keeping versions aligned, and managing configuration changes.

Balancing agentic autonomy and control

In the enterprise, sanctioned MCP use is proving to be incredibly powerful. Just take the case of Workato, which experienced a 700% increase in Claude chats from internal employees over a 60-day period when it turned on enterprise MCP features. Support engineers, financial analysts, sales leads, and others are building new workflows that grow Workato’s business in tangible ways, much in part thanks to MCP.

Getting those results, however, requires balancing agentic autonomy with control. That’s where an MCP registry can shine. For an enterprise, the quality of an MCP registry doesn’t just depend on listing every MCP server in a directory. It hinges on trust, safety, and smart controls — especially to prevent leaking data from chat streams across inter-organizational agent workflows, for instance.

As such, enterprises going “all in” on MCP should seriously consider MCP registries as a core infrastructure, with all the standard architectural enterprise bells and whistles. “It should be treated like any other serious piece of software,” says Alareqi. “That means strong versioning, life-cycle management, and observability.”

Kotlin 2.3.20 has become the latest version of the JetBrains-built language, featuring an interoperability mode for C or Objective-C libraries and name-based destructuring declarations for property names. Developers also can leverage Kotlin interfaces on JavaScript and TypeScript.

The update to the Java rival language was introduced March 16. Instructions for getting started with the language can be found on the Kotlin website. With the Kotlin Native technology in Version 2.3.0, for compiling Kotlin code to native binaries, developers can try the now-experimental interoperability mode for Objective-C and C libraries. This capability is geared to developers who use C or Objective-C libraries in Kotlin Multiplatform (KMP) libraries or applications. In general, Kotlin Native enables importing C and Objective-C libraries into Kotlin. However, for KMP libraries, this functionality is currently affected by the KMP compatibility issues with older compiler versions. Thus, if a KMP library compiled with one Kotlin version is published, importing C or Objective-C libraries might make it impossible to use that Kotlin library in projects with an earlier Kotlin version. To address this and other issues, the Kotlin team has been revising the interoperability mechanism. Starting with Kotlin 2.3.20, developers can try the new mode through a compiler option.

Also Kotlin 2.3.20 introduces name-based destructuring declarations that match variables to property names instead of relying on position-based componentN() functions. Previously, destructuring declarations used position-based destructuring, JetBrains said.

The update lifts the limitation on implementing Kotlin interfaces on the JavaScript and TypeScript sides, JetBrains said. Previously, it only was possible to export Kotlin interfaces to TypeScript as TypeScript interfaces; implementing them from TypeScript was forbidden. Additionally, starting with Kotlin 2.3.20, Kotlin/JS supports the SWC Rust-based compilation platform. This helps with transpiling newer versions of JavaScript and TypeScript code into older and more compatible JavaScript code.

Kotlin 2.3.20 follows the December 2025 release of Kotlin 2.3.0 and the February release of Kotlin 2.3.10. Elsewhere in Kotlin 2.3.20:

• For Java interoperability, the compiler now recognizes the Vert.x @Nullableannotation for nullability checks. This release also adds support for the Java @Unmodifiable and @UnmodifiableView annotations to treat annotated collections as read-only in Kotlin.
• It is easier to set up Kotlin in Maven build tool projects. Now, Kotlin supports the automatic configuration of source roots and Kotlin’s standard library.
• Kotlin 2.3.20 is fully compatible with Gradle build tool Versions 7.6.3 through 9.3.0. Developers also can use Gradle versions up to the latest Gradle release. Developers should be aware that doing so may result in deprecation warnings, and some new Gradle features might not work.
• The Lombok compiler plug-in for generation and use of Java Lombok declarations has been promoted to alpha status. Plans call for making this functionality production-ready, but it is still under development.
• The Map.Entry.copy() extension function is introduced for creating an immutable copy of a Map.Entry. This function allows for reusing entries obtained from Map.entries after modifying the map by copying them first.

AI models cost a lot more to develop than you may think. AI research company Epoch AI has set out all the costs of building a new AI model — and explaining why AI companies are so concerned about perceived threats to their intellectual property.

It has looked into this before: Last year, it estimated that of OpenAI’s $5 billion expenditure on R&D, only about 10 percent went on the final training runs, with the majority going on scaling, synthetic data generation, and basic research.

At the time, Epoch was unsure whether this was a peculiarity of OpenAI but now two Chinese companies, MiniMax and Z.ai, have also disclosed their R&D compute spending, and Epoch has found that, despite the differences in company size, final training runs are only a small part of the Chinese companies’ R&D expenditure too.

Epoch set out more detail about the issue. It said that if “most of the spending is exploration rather than execution, then a competitor who learns what works from the frontier could replicate the results for a fraction of the original cost.”

This has been a concern of US AI companies for some time.  Google has already expressed concerns about intellectual property theft. And Anthropic has fingered MiniMax as a company that has sought to extract Claude’s capabilities to enhance its own offerings. It’s clear that any business looking to develop AI models is going to be committing to spend huge sums of money: The training is just a small part of it.

OpenAI has introduced a plugin system for Codex, its AI-powered software engineering platform, giving enterprise IT teams a way to package coding workflows, application integrations, and external tool configurations into versioned, installable bundles that can be distributed or blocked across development organizations.

“We’re rolling out plugins in Codex,” OpenAI Developers, the company’s official developer account, posted on X.  “Codex now works seamlessly out of the box with the most important tools builders already use, like Slack, Figma, Notion, Gmail, and more.”

Plugins are “installable bundles for reusable Codex workflows” that “make it easier to share the same setup across projects or teams,” an OpenAI developer portal documentation noted. Each bundle can contain skills, which the documentation describes as prompts that the Codex agent can discover and execute, along with optional application integrations and Model Context Protocol server configurations that give the agent access to remote tools or shared context, it added.

A governance layer for agentic AI

How those bundles are distributed and governed is controlled through a separate policy layer, the documentation said.

Organizations can define plugin catalogs, called marketplaces, in JSON files scoped either to a repository or to an individual developer’s environment. Each plugin entry carries an installation policy with values including “INSTALLED_BY_DEFAULT,” “AVAILABLE,” and “NOT_AVAILABLE,” giving administrators the ability to push, restrict, or block plugins across the developer workforce, the document added. Authentication behavior is configurable at the policy level as well.

The plugin feature is the latest in a run of enterprise-focused additions to Codex since OpenAI announced the platform’s general availability in October 2025, when it said Cisco had reported pull request review times falling by as much as 50% after deployment. Admin tooling released at the same time gave ChatGPT Business, Edu, and Enterprise customers environment controls, usage analytics dashboards, and managed configuration options for the Codex CLI and IDE extension.

“Centralized control over which plugins are permitted, blocked, or deployed by default directly addresses concerns around security, compliance, and operational consistency,” said Charlie Dai, VP and principal analyst at Forrester. “It aligns AI agents with existing IT governance models rather than bypassing them.”

Adoption will be gradual, Dai said. “While technical tooling is advancing quickly, most enterprises will adopt this incrementally, led by platform engineering and developer productivity teams,” he said.

Agent behavior as managed infrastructure

Beyond the pace of adoption, Dai said the plugin system signals a broader shift in how enterprises are expected to manage AI-assisted development.

“By encapsulating standards, workflows, and tool access into versioned artifacts, organizations elevate AI-assisted development from ad hoc usage to managed infrastructure,” he said.

That distinguishes Codex from its main rivals. GitHub Copilot Extensions, which reached general availability in early 2025, lets developers invoke third-party tools from Copilot Chat inside Visual Studio Code, JetBrains IDEs, and GitHub.com, with a public marketplace hosting extensions from vendors including Docker, Sentry, and Perplexity. The emphasis is on contextual tool access during chat sessions rather than governing agent behavior at scale.

Cursor, another rival, launched its own plugin marketplace in February. The company expanded it this month, adding more than 30 integrations from partners including Atlassian, Datadog, and GitLab, according to Cursor’s changelog. Teams and Enterprise administrators can also create private marketplaces for controlled distribution.

Anthropic has moved in a similar direction, introducing workflow automation plugins for its Claude Cowork platform earlier this year.

“Compared with GitHub Copilot or Cursor, OpenAI is extending beyond policy enforcement into behavioral standardization,” Dai said. “Competitors focus primarily on permissions and guardrails; Codex begins to formalize execution patterns at scale.”

The missing third-party ecosystem

That behavioral standardization, however, has a notable constraint for now.

OpenAI has not opened self-serve publishing to its official plugin directory. “Adding plugins to the official Plugin Directory is coming soon,” the documentation said. “Self-serve plugin publishing and management are coming soon.” Organizations are limited for now to private marketplaces scoped to a repository or to an individual developer’s environment.

On the other hand, GitHub’s marketplace has been open to third-party builders since early 2025. Cursor’s marketplace already lists more than 30 external partners. OpenAI’s directory so far contains only plugins curated by the company itself.

“Long-term platform stickiness will depend on a curated third-party ecosystem that expands capability breadth and accelerates innovation,” Dai said. “Mature enterprises will expect audited, interoperable plugins for domain-specific tooling and regulated workflows. Without this external ecosystem, Codex risks limited extensibility beyond core engineering use cases.”