Brex built its AI agent policy by watching what agents actually do, not by writing rules first

Brex built an internal open-source proxy called CrabTrap to monitor HTTP/HTTPS traffic from AI agents using real credentials such as API keys and OAuth tokens. The system intercepts requests, checks policy rules, and uses an LLM-as-a-judge approach to approve or deny agent actions based on observed behavior rather than predefined SDK rules alone.

The approach reflects a shift from static guardrails toward centralized agent governance at the network layer, where enterprises can see and control what agents actually do. If adopted more broadly, this could become a key security pattern for scaling autonomous agents without stripping away their usefulness.