When you hand a job to an AI employee, you are handing it data. That demand a clear posture on isolation. Here is the runbook we follow before any worker goes live.
One container, one worker
Each AI employee runs in its own container. There is no shared state, no cross-tenant access, and no worker that can reach another worker’s scope. If one misbehaves, it is contained — not a fleet-wide incident.
A private network, not the open internet
Workers sit on an isolated network. They reach the tools you authorise — and only those. The control plane manages the fleet; workers never manage each other or the platform itself.
The verification gate
Before any agent goes live it must pass an isolation check covering socket access, network scope, and privilege. If the gate does not pass, the worker does not ship — no exceptions for convenience.
Backup is part of safety
Every worker is backed up with copy/replace discipline, so the workforce can be restored to any point after a change. Safety is not just preventing a breach; it is guaranteeing you can always get back to a known-good state.
A workforce you cannot isolate and restore is a liability. The runbook exists so that never happens.