How Postman runs Agent Mode for 40 million developers on Amazon Bedrock
aws ·
Postman has successfully deployed Agent Mode for its 40 million developers using Amazon Bedrock, addressing the complex engineering challenges of integrating AI into a mature product. The team discovered that managing tool sprawl was critical, implementing dynamic tool selection to limit the model's view to only relevant tools and reducing errors associated with large toolsets. To improve efficiency, Postman shifted from creating numerous single-purpose tools to providing schema-based read access, allowing the agent to query structured data directly via SQL-like commands. Context management emerged as the primary bottleneck, leading to the development of purpose-built context handlers that distill entity information rather than serializing interface-specific data models. On the infrastructure side, Amazon Bedrock provides essential capabilities such as model flexibility across the Claude family, cross-Region inference for handling traffic bursts, and tiered prompt caching to optimize latency and costs. Enterprise security is maintained through geographic inference profiles and zero data retention settings, ensuring compliance with strict data residency requirements. These architectural patterns demonstrate how teams can move production agents beyond prototypes by treating context and tool exposure as scarce resources. Ultimately, the solution highlights the importance of decoupling agent actions from UI state and maintaining a robust Retrieval Augmented Generation knowledge base.
aws