Brex has developed an AI agent governance framework built on observed agent behavior rather than predefined rule sets, using a centralized network control plane and structured agent mesh to enforce policies across financial and operational workflows. Rather than authoring governance rules in advance, Brex bootstrapped its policy creation from real network traffic, tool usage, and decision patterns captured in production environments. Agents initially operated in shadow mode, collecting behavioral data without influencing live decisions. A policy builder loop then translated those observed actions into natural-language policies, evolving governance away from SDK-level permissions and static guardrails toward controls grounded in operational evidence.
Brex builds AI governance from observed behavior, not assumptions — letting real agent actions write the rules.
The network control plane, called CrabTrap, functions as an HTTP/HTTPS proxy positioned between agents and the external internet. It intercepts all agent network traffic and evaluates each request against security and policy criteria. Deterministic static rules handle the majority of recognized patterns, while an LLM-as-judge component handles unfamiliar endpoints or unusual request shapes. That LLM judgment is reportedly required for fewer than 3% of requests in mature agent deployments, concentrating oversight on long-tail risks rather than routine traffic. The design enables centralized enforcement while continuously learning from agent behavior across systems.
Brex also built an Agent Mesh composed of narrow, role-specific agents that communicate over a shared message stream. A config layer defines each agent’s role, model, tools, and subscriptions, providing consistent governance across the network. A MessageStream component logs every message, tool call, and state change, producing replayable pipelines and full audit trails. A Clock component enforces deterministic event ordering, supporting both reliable execution and post-hoc analysis. Governance is embedded directly into the system architecture rather than added as an external layer, keeping reasoning, privacy boundaries, and policy enforcement inspectable by design.
Every agent approval or rejection must include a clear reasoning trail to support downstream review. Agents exchange structured claims, supporting evidence, and confidence levels, enabling deterministic execution and meaningful auditing. Dedicated audit agents review decisions to verify adherence to accuracy and behavioral policies across workflows. Human interventions are captured as labeled data points that improve future automation and refine policy calibration over time. Audit frameworks use multi-turn evaluations and simulated end-user objectives to detect regressions and quality issues in agent behavior before they affect production outcomes.
The overall governance approach reflects a broader design philosophy at Brex: policies should emerge from evidence rather than assumptions. By grounding controls in actual agent behavior, building inspection into the infrastructure itself, and maintaining continuous human-in-the-loop feedback, the framework is intended to remain accurate as agent capabilities and use cases evolve. Today’s AI agents can plan complex workflows and access enterprise applications, making observation-driven governance essential for keeping pace with dynamic and difficult-to-predict behaviors. Adversarial testing further stress-tests policy boundaries, ensuring that governance keeps pace with changes in how agents operate across financial and operational domains.
