Brex assumes its AI agents could do anything — so it watches the network, not the code

Brex CEO Pedro Franceschi outlined a security-first approach to deploying AI agents in enterprise environments, rejecting the industry's reliance on code inspection as the primary control mechanism. Instead of monitoring what agents do internally, Brex watches network activity to detect suspicious behavior before it causes damage.

The company treats AI agents as black boxes with unpredictable outputs. Since traditional code auditing cannot guarantee what a sufficiently advanced AI system will do next, Brex shifted focus to external monitoring. The strategy assumes agents could execute any action and thus requires observation of their actual network requests, API calls, and data access patterns in real time.

Franceschi criticized the term "agent" itself as vague Silicon Valley jargon that obscures the real problem: how to safely give autonomous systems access to business-critical systems. He advocated reframing the conversation around "virtual employees" that collaborate with human workers rather than replacing them.

This distinction matters for deployment. A virtual employee model requires explicit permission boundaries and human oversight rather than assuming agents operate within narrowly defined parameters. Brex's network monitoring approach reflects this reality. If an AI system can theoretically do anything, you cannot secure it by reading its source code. You secure it by controlling what it touches and monitoring what actually happens.

The approach carries real operational weight. Brex is already deploying open-source tools like OpenClaw into production, which means this security philosophy isn't theoretical. Network-based detection catches anomalies in real time. If an agent suddenly requests access to sensitive customer data or initiates unexpected external communications, the system can block or alert before harm occurs.

This represents a maturation in enterprise AI thinking. Rather than betting that careful prompt engineering and architectural constraints will constrain agent behavior, companies like Brex assume worst-case scenarios and build infrastructure accordingly. The shift from "trust the code" to "monitor the