Postman shares how it runs Agent Mode for 40 million developers

Summary

Postman and AWS published an official blog post on the engineering behind Agent Mode, Postman's AI-native way of working across API testing, documentation, discovery, and implementation. The team treats context, not capability, as the main bottleneck: a vector search narrows more than 170 tools to roughly 15 per request, actions that change application state require user approval, and Bedrock Guardrails redact personal information before it reaches the model. Postman's internal testing also found that tool-selection errors grew once the visible toolset exceeded about 40 tools. The 40-million-developer scale and internal figures are company-reported.

Why it matters for our work

The post shows concretely what separates a demo agent from a production one. The observation that exposing every tool increases mistakes offers a practical lesson for anyone building human-AI collaboration tools.

Translated from the Korean original. Summaries may be translated and edited. Commentary reflects our perspective; forecasts remain the source’s views.