Postman’s new Agent Mode is an AI‑native portal that lets developers work across API testing, documentation, discovery, and implementation without leaving the Postman interface. Built on Amazon Bedrock, the system leverages Bedrock’s managed foundation models, cross‑region inference, and multi‑tier prompt caching to serve a global developer base of 40 million users. To keep the agent focused and safe, Postman dynamically scopes the tool set to roughly 15 relevant tools per task, using a vector database of tool embeddings to filter a larger catalog of 170 tools. The agent also relies on schema‑aware read access to structured data, such as ClickHouse tables, allowing it to generate complex queries instead of requiring a separate tool for each question. Human approval is required for any action that modifies application state, and Bedrock Guardrails redact personally identifiable information before it reaches the underlying large language model. Context is supplied in two layers: a broad, shallow background context automatically minified for the prompt, and a deep, focused context chosen by the user and routed through dedicated handlers that distill each entity into the information the agent needs. This design reduces failures caused by missing or incorrect context, which proved more problematic than missing tool capabilities. Together, these architectural patterns enable Postman to run a production‑grade AI agent at scale while maintaining control over data exposure, model selection, and throughput, illustrating a practical path for other teams to move beyond prototype agents into production deployments.
How Postman runs Agent Mode for 40 million developers on Amazon Bedrock
Postman has launched Agent Mode, an AI‑native interface that integrates with Amazon Bedrock to support API testing, documentation, discovery, and implementation for its 40 million‑developer community. The solution uses dynamic tool scoping, schema‑aware data access, and context‑centric design to keep the agent focused and reduce unintended actions. Human approval is required for state‑changing operations, and Bedrock Guardrails redact personal data before LLM inference. The architecture scales globally without running its own model‑serving infrastructure, leveraging Bedrock’s cross‑region inference and multi‑tier prompt caching.
Why it matters
The approach demonstrates how large‑scale AI agents can be safely integrated into mature products, balancing automation with human oversight and data privacy—critical for broader AI adoption in enterprise software.