Brex Reinvents AI Agent Policy with Network‑Level Enforcement, Not Pre‑Written Rules

Brex has turned the conventional approach to AI agent policy on its head by letting real‑world traffic dictate rules rather than pre‑written code.

Network‑Layer AI Agent Policy Enforcement

The open‑source OpenClaw framework is popular, yet enterprises stumble over credential leakage and unchecked API calls. To close the gap Brex built CrabTrap, an HTTP/HTTPS proxy that intercepts every outbound request, checks static rules and, when needed, hands the decision to an LLM‑as‑a‑judge. ‘Every request an agent makes is an opportunity to intercept, reason about, and make a policy decision,’

Pedro Franceschi, Brex co‑founder and CEO, told Reuters

By positioning enforcement at the transport layer, the system stays framework‑agnostic and requires only an HTTP_PROXY setting. The LLM judge fires on the long tail – roughly ↑ 3% of calls – keeping latency negligible (↓ 1ms overhead). Policy generation is automated: a shadow‑mode agent streams historic traffic, the builder drafts a natural‑language policy, and an evaluator validates changes before rollout. Audits land in PostgreSQL, enabling dashboards that surface denial patterns and trigger human‑in‑the‑loop reviews. The result, according to Franceschi, is ‘organizational confidence’ to expand autonomous agents across finance, HR and ops. Open‑source contributors are invited to add SSO, RBAC and escalation workflows, turning CrabTrap from a hard fence into a managed permission platform. The broader lesson for security teams: don’t wait for industry standards; build layered controls now. For further reading see Bloomberg.