Skip to content
worth noting AI agents

Postman and AWS described the architecture of Agent Mode on Amazon Bedrock

only one source so far

Agent Mode in Postman runs on Amazon Bedrock. The case study describes task-based tool selection, reading data using schemas and user approval of changes. According to Postman, errors increased when more than approximately 40 tools were made available.

Postman and AWS published a case study on the operation of Agent Mode, which helps with API testing, documentation, search and implementation in Postman. The feature runs on Amazon Bedrock. According to Postman, this solution makes it possible to handle variable workloads without maintaining its own infrastructure for running models, while retaining the ability to choose models. The article cites a Postman developer community of 40 million people; this is not a documented count of Agent Mode users.

According to internal testing by Postman, tool selection errors increased when the model could see more than approximately 40 tools. The current architecture therefore makes only the tools relevant to the task available. For API Catalog, the company combined several narrowly focused views into a single query tool: the agent uses ClickHouse table schemas to generate queries against structured data. According to the company, missing or incomplete context caused more failures than missing capabilities.

Actions that change the application state require user approval. Agent Mode can already send requests in the background without an open tab, but approval is still required. Data retention settings depend on the model used. Enterprise administrators can use Amazon Bedrock Guardrails to enable the removal of personal data before it is passed to the underlying model. Details are available in the source article.

What changed

Why it matters

Teams integrating an agent into an established application gain a concrete example for designing data access and tool selection. In the experience of Postman, larger or newer models reduced errors when a large number of tools were available, but did not eliminate them. The results therefore support validating context and tool interfaces when designing a solution; the threshold of approximately 40 tools comes from testing by this company.

Two audiences, two different impacts

What this means

01

For individuals

A developer can send requests through Agent Mode without an open tab. For actions that change the application state, approving the proposed action remains part of the workflow.

What to do When using Agent Mode, review the proposed changes before giving approval.
More practical updates →
02

For a business

For development teams, the case study shows how to limit the number of tools made available to a model: structured data can be read using a single query tool instead of creating a separate tool for each question.

Development
What to decide Test making tools available based on the current task using your own tasks.
More business impacts →
Agent Mode Amazon Bedrock Amazon Bedrock Guardrails Postman Prompt caching

Check the original

Event sources

only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.

1
AWS Machine Learning Blog primary source · first detected How Postman runs Agent Mode for 40 million developers on Amazon Bedrock