Postman says Agent Mode uses Amazon Bedrock and task-specific tools, context handling and inference controls to support its developer community at production scale.
Postman has outlined how it built Agent Mode, an AI interface for working across API testing, documentation, discovery and implementation. The company says its product serves a community of 40 million developers; the system runs on Amazon Bedrock, which provides managed access to foundation models without Postman operating its own model-serving infrastructure.
A central design challenge was keeping the agent’s available actions focused. In Postman’s testing, tool-selection errors increased when more than approximately 40 tools were visible to the model. Its architecture can search a catalog of more than 170 tools and narrow the selection to approximately 15 relevant tools for a task. The agent also uses schema-aware reads to query structured data, reducing the need to create a separate tool for every question.
Postman says missing or poorly prepared context caused more failures than missing capabilities. It built dedicated handlers to extract useful details about selected product objects rather than passing along data shaped primarily for the application interface. Agent Mode also combines its tools and instructions with a retrieval-augmented knowledge base that selects relevant product documentation for a request.
The Bedrock setup supports routing work to different Anthropic Claude models, cross-Region inference profiles and prompt caching. Postman describes using five-minute and one-hour cache checkpoints for context that changes at different rates. It has also configured zero data retention for supported models; availability depends on the model. Amazon Bedrock Guardrails can redact personally identifiable information before it reaches the model, and enterprise administrators can enable that setting.
Agent Mode requires user approval for actions that change application state. Postman says it can send requests in the background without an open tab, but approval is still required. The production implementation is proprietary and is not available as a public sample repository.