You Can't Monitor AI Agent Traffic with the Network You Own

Autonomous agents now open connections on their own: calling model and inference providers, calling other agents, and reaching services no human requested. This traffic is exploding, and it is concentrated where the stakes are highest.

An IP address is not an agent

The controls most enterprises own were built for hosts and humans. An IP address is not an agent. A login is not an agent. Dozens of ephemeral agents share one host and one address, so the flow tools that monitor network traffic by IP and port can describe the lanes but cannot name the vehicles. When an auditor asks which agent reached which provider last Tuesday, the stack you own has no answer.

How to monitor AI agent traffic and stay in control

To monitor AI agent traffic in a way that survives an audit, three things have to hold:

• Identity per agent, from the wire. Attribution has to come from the kernel and network level, not from telemetry the agent emits about itself, and not from the address it happens to share.
• Declarations from the operator. You declare what each agent may reach; the declaration is the baseline, not a model's guess about what looks anomalous.
• Deviations, not dashboards. When behavior contradicts the declaration, a typed deviation fires: declared agent, unlisted provider. Your SIEM and your team stay the judgment.

Maya does exactly this: it identifies every agent from the wire, payload-blind, in your VPC, with no SDK and no sidecar, and holds each one to the scope you declared.

We will use this space for what we learn watching real fleets.