A fence round the most privileged seat
AgentOS runs coding agents in containers, and the team-lead seat is the most privileged of them. It directs other agents and uses the estate's governance tools. My view was that a seat like that should reach its model vendor and nothing else, so I built deny-by-default network confinement for it. The seat's container sits on an internal network with no route out. Its one exit is an allow-listing proxy, and the proxy carries one thing: an encrypted connection to the seat's vendor endpoints.
I treat the agent's network as an attack surface and start from deny-all with one named exit.
The usual bypasses, closed
An allow-list has well-known ways around it, and each is closed explicitly.
The seat cannot reconfigure the proxy, and when the proxy is full it refuses new tunnels. The proxy holds no credentials, so confinement adds no secret to protect.
A rollout that proves the blocks
Rollout proves the blocks, in both directions, before any seat is trusted to them.
Network confinement sits beside credential isolation: the seat's logins are isolated from other seats, and its traffic reaches only its vendor. The firewall, container networking and TLS are standard parts; the policy, the proxy and the proof-driven rollout and preflight are my design.