Every agent task gets its own disposable environment. The broker decides scope, the claim creates it, and the only way out of the sandbox is the governed gateway.
Requirement
Design and run the execution environments agents work in: sandboxed, reproducible, permissioned, disposable.
Design choices
Identity per session, not per agent type: 30 minute TTL, scope derived from the task.
No route out except the gateway; no standing secrets inside the sandbox.
Namespace carries a TTL and finalizer; rebuildable from an image digest plus the task spec.
Evidence and gap
Built in Hangar: the claim-a-thing, get-a-governed-environment pattern (Crossplane XRs) and the Tower policy of no standing pod exec. Proposal only: gVisor or Kata runtime and per-session identity minting.