How Postman runs Agent Mode for 40 million developers on Amazon Bedrock

Building a demo agent and running one inside a mature developer platform are very different jobs. Postman says Agent Mode had to work across testing, documentation, discovery, and implementation for a user base of 40 million developers, and that the real challenge was not just model quality or prompt design. The product already had years of interface habits baked into it, so the team had to teach an agent to reason over data instead of poking around tabs and sidebars like a human with a mouse.

That led Postman to a few blunt lessons. First: tool sprawl hurts. The team found that once the visible toolset grew beyond about 40 tools, selection mistakes rose. The answer was to stop handing the model everything and instead narrow the tools per task, with a root agent picking from more than 170 tools and passing only about 15 relevant ones to a context-isolated sub-agent. Postman also says some of its client APIs were tied too tightly to UI state, so it is actively decoupling tools from tabs. Agent Mode can now send requests in the background, though user approval is still required for state-changing actions.

Second: if the data is structured, let the model query it. For products like the API Catalog, Postman consolidated narrow read tools into schema-aware access to a query engine, using ClickHouse tables to answer questions about service uptime, test results, endpoint response times, and more. That shifts the work from inventing a separate tool for every question to modeling the data well once. The company argues that this gives the agent a much better scaling curve than an ever-growing pile of one-off reads.

The biggest surprise was context, not capability. Postman expected missing tools to be the main failure mode, but found that missing or incomplete context caused more breakage. It built dedicated context handlers because serializing the existing interface data model was good for rendering and transfer, not for reasoning. And even that ran into a new problem: the context window got crowded by open-ended user data such as request descriptions, OpenAPI specs, and payloads. In other words, the agent didn’t just need the right tools; it needed the right crumbs.

Amazon Bedrock is what lets Postman run this at production scale without running its own model-serving stack. The setup gives it model flexibility across the Claude family, geographic and global cross-Region inference for bursty traffic, model-dependent zero data retention, multi-tier prompt caching, and Bedrock Guardrails to redact personally identifiable information before it reaches the model. The larger point is pretty clear: production agents are less about flashy demos and more about relentless control over tools, context, and where the data goes.