Agent Heartbeat
A periodic signal emitted by a running agent to its watchdog or orchestrator to indicate it is alive and making progress, enabling detection of stuck, crashed, or infinitely looping agents.
Heartbeats are a standard pattern in distributed systems for detecting failures, applied here to AI agents. The agent emits a heartbeat—a lightweight status update (agent ID, task ID, current step, tokens consumed so far)—at the end of each loop iteration or on a fixed timer. The watchdog expects heartbeats at regular intervals; a missed heartbeat triggers an alert and optionally a kill signal.
Heartbeat design should include progress information, not just liveness. An agent that is running but not making progress (e.g., retrying the same failed tool call in a tight loop) is just as problematic as a crashed agent. Including a "step count" or "cost delta" in the heartbeat allows the watchdog to detect no-progress loops.
In Paperclip's watchdog implementation (`watchdog/watchdog.mjs`), agents emit heartbeats via task status updates. The watchdog monitors task update timestamps and triggers a circuit breaker if a task has been in a non-terminal state for longer than its configured maximum duration.