Postman runs Agent Mode for 40 million developers on Amazon Bedrock

The team found that the hardest problems were not model quality but integrating an agent into a mature product. They limited the tool catalog to about 40 atomic tools, dynamically scoped them per task, and built schema‑aware read handlers to reduce tool sprawl. Context budgeting proved critical; dedicated handlers distill only the data the model needs, while prompt caching (one‑hour and five‑minute checkpoints) cuts latency and cost. Human oversight and Bedrock Guardrails protect against unintended actions and data exposure.

Postman’s architecture separates client‑side tools, generic agent instructions, and a retrieval‑augmented knowledge base. By leveraging Bedrock’s model flexibility and geographic inference profiles, the service can handle bursty developer traffic without running its own model‑serving infrastructure, while keeping data residency controls for enterprise customers.