Postman brings Agent Mode to 40 million developers on Amazon Bedrock

Postman and AWS have explained how Agent Mode works. It is an AI agent inside the platform for designing, testing, and documenting APIs. The product has about 40 million developers, and the agent calls models through Amazon Bedrock. The core argument of the write-up: building an agent that looks good in a demo is relatively easy. Making it work reliably for an audience this size is hard, and the difficulty lies in the architecture around the model.

The three principles behind Agent Mode

The authors describe three patterns. The first is control over tool sprawl. A platform like Postman offers a great many actions: create a collection, run a request, edit an environment, generate tests, update documentation. If you hand the agent all of them as a flat list, the tool descriptions take up a noticeable share of the context. The model starts to confuse similar functions and more often calls the wrong one. So the toolset has to be deliberately narrowed and structured.

The second principle is reading through schemas. Instead of dumping everything known about a project into the prompt, the agent consults a structured description of the API and pulls only the piece it needs. This is a natural move in the API world: specifications already have a formal structure, and it can serve as a map of the terrain.

The third principle is the most down-to-earth. Context is the real bottleneck. Not the model's intelligence and not token speed, but what exactly goes into the window and how much of it. For an agent, this affects accuracy, latency, and the cost of every request at once.

Why Amazon Bedrock

In this story, Bedrock serves as the infrastructure layer. It is a managed AWS service that gives access to various foundation models through a single API, with no need to run your own GPU clusters. For a product with tens of millions of users, this means scaling for peak load, the security mechanisms enterprise customers are used to, and the ability to swap models without rewriting the whole application.

An important caveat: the published description contains no data on which models Postman uses, what the latencies are, what each request costs, or what share of tasks complete successfully. Without these numbers, the architectural principles remain the team's claim rather than a measured result.

What the announcement tells us

• Product: Agent Mode in Postman, an agent for working with APIs.
• Audience: about 40 million developers.
• Inference platform: Amazon Bedrock.
• Key patterns: limiting the toolset, reading through schemas, context management.
• Not disclosed: the models used, latency, cost per request, quality metrics.

What to take into your own stack

If you are building your own agent, the material offers three practical takeaways.

First: audit your tools. Count how many tokens their descriptions consume before work even begins. If the set has grown past a few dozen functions, split it by scenario and connect only the ones relevant to the current task.

Second: give your agent a data-reading layer on top of formal schemas, whether OpenAPI, JSON Schema, or a database schema. An agent that requests the fragment it needs by structure uses less context and hallucinates less often than one with documentation dumped on it.

Third: measure context the same way you measure latency and cost. The input size at each step, the share of useful tokens, and the number of repeated requests for the same data tell you more about an agent's quality than a subjective "it answers fine."

The main takeaway for the market: competition among agents is gradually shifting from the choice of model to the engineering around it. You can swap a model on Bedrock in an evening. Well-designed handling of tools and context cannot be replaced that quickly.