Postman launches Agent Mode on Amazon Bedrock

Postman’s new Agent Mode provides an AI‑native way to work across API testing, documentation, discovery and implementation. The feature runs on Amazon Bedrock, which supplies managed access to foundation models, geographic inference routing, model‑dependent zero‑data‑retention and multi‑tier prompt caching, allowing Postman to handle bursty, latency‑sensitive developer traffic without operating its own model‑serving stack.

Engineering the product revealed three key challenges. A large catalog of atomic tools (>170) caused selection errors and slow interactions; Postman now narrows the tool set to roughly 15 relevant tools per request using a vector‑based embedding lookup. By exposing schema‑aware read access to ClickHouse tables, the agent can generate complex queries, reducing the need for many single‑purpose tools. Context proved to be the primary bottleneck, leading the team to build purpose‑shaped context handlers and treat the context window as a scarce budget. Human oversight remains mandatory for any state‑changing action, and Bedrock Guardrails redact personally identifiable information before it reaches the underlying LLM.

For AI developers, the post outlines practical patterns for scaling production agents: dynamic tool scoping, schema‑driven data access, disciplined context management, and built‑in safety controls. These lessons can help teams move beyond prototypes toward reliable, enterprise‑grade AI agents.