How Postman Structures Agent Mode on Amazon Bedrock

Postman's Agent Mode runs on Amazon Bedrock — but the interesting engineering isn't the model, it's everything around it. Postman and AWS describe three connected decisions that make the agent work: narrowing the tools shown to each task, preparing context through dedicated paths, and keeping human approval in the execution flow.

From 170+ tools down to ~15 per task

Postman reports that tool-selection errors climbed in its testing once the visible toolset exceeded roughly 40 tools. The response: task-based retrieval. Rather than exposing the entire 170+ tool catalog, Agent Mode searches for relevant capabilities and narrows the set to about 15 for a given task — then hands them to a context-isolated sub-agent. The retrieval step becomes consequential: it decides which capabilities the executing agent receives.

For structured data, Postman also describes schema-aware read access to ClickHouse tables — querying structured information instead of minting a narrow tool for every question.

Two paths for context

Postman identifies missing or incomplete context as a major bottleneck. Background context is gathered automatically from the workspace and reduced before entering the prompt; focused context comes from user selections or mentions and passes through handlers tailored to the relevant Postman object. The trade-off is explicit: reducing context limits what reaches the model, so workspace access doesn't guarantee every relevant detail survives prompt preparation.

Inference routing, caching, and redaction

Agent Mode reaches Claude models through Bedrock's inference APIs, using cross-Region inference profiles to route workloads. Prompt caching uses a one-hour checkpoint for stable content and a five-minute checkpoint for variable context — with model-dependent support limits, not a blanket guarantee. On the privacy side, Postman uses Bedrock Guardrails to redact PII before it reaches the model, configurable by enterprise admins.

Approval stays in the loop

Agent Mode requires user approval before actions that modify application state — and background request sending doesn't waive that. Redaction and approval address different things: information passed to the model versus changes made through the application.

The useful walkthrough, the open questions

The case study is best read as an architectural walkthrough — how Postman scopes tools, prepares context, routes inference, and retains human control. It documents the mechanisms without supplying measured task-success rates or an independent safety evaluation. One clarification from the coverage: the "40 million developers" headline refers to Postman's developer community, not Agent Mode users.

Go deeper

This is a condensed overview. The full analysis — with source citations, the evidence caveats, and the architecture diagrams described — is on Agentic Affairs:

👉 How Postman Structures Agent Mode on Amazon Bedrock

---

Originally published on Agentic Affairs (https://agentic-affairs.com)