Agent shutdown control must reach every active execution path, including background jobs, subagents, queues, scheduled work, retries, remote browsers, and external tool sessions. Stopping a visible chat is insufficient if delegated work continues elsewhere.
A reliable shutdown revokes or disables credentials, interrupts execution safely, records what stopped, preserves diagnostic evidence, prevents automatic restart, and reports any actions that could not be cancelled or rolled back.
ELI5
An agent shutdown control is a dependable way to stop an AI agent and the work it has delegated. It must stop background tasks and access, not just close the screen where the agent was visible.
For example, pressing stop can cancel queued jobs, end remote sessions, and revoke the agent's temporary credentials. The system should report any external action that had already completed and could not be undone.
