Fail-closed agent orchestration

An agent can finish a task after an operator has stopped the run. If that late result is accepted, work the operator meant to halt may still affect what happens next. Skein is an early-stage, open-source tool for coordinating local agent work. It uses an epoch fence, a marker that changes when a run is stopped, and rejects completions from an older epoch rather than just labeling those tasks canceled. The repository has 259 stars. Skein’s GitHub page warns that teams should test its stop behavior on their own workloads before relying on it operationally.

That distinction matters if you’re running agents that can launch commands or handle multi-step tasks. A stop button is only useful if work already in flight can’t quietly come back and be treated as current. Skein says its protection covers scheduling in one local run and termination of tracked local process trees. It doesn’t cover remote work, side effects that have already crossed into external systems, or checking whether an agent’s output is correct. So this is a control over task execution, not a guarantee that an agent can’t cause harm.

For an owner, the practical question is whether your own stop process reliably prevents stale work from being accepted, and what can still happen outside that boundary. That’s the kind of safeguard to track in an AI command centre, alongside where your agents run and what they’re allowed to do.