A pattern where AI agents proactively wake up on a schedule (every 4, 8, or 12 hours) to re-read their context, check their task queue, and continue work — without a human prompt triggering them. Popularised by OpenClaw, now a core mechanism in Paperclip.
Key points
- Each heartbeat starts with a fresh context window — as if the agent were just born. The agent must immediately re-read its instruction files (
agents.md,heartbeat.md,soul.md,tools.md) to re-establish identity and priorities before doing anything else [src-001].
- This creates a specific discipline requirement: agent instruction files must be comprehensive and current, because the agent has no memory of prior heartbeats except what’s in those files and the ticketing system.
- The heartbeat pattern is what makes agents feel proactive — they appear to keep working on their own, and can take initiative (flag blockers, request approvals, ask for missing info) without waiting for the human to prompt them [src-001].
- Heartbeat frequency is a trade-off: more frequent = faster iteration but higher cost and more context thrashing. Less frequent = cheaper but slower responsiveness. Paperclip defaults to 4/8/12 hours as configurable presets [src-001].
- Contrast with event-driven triggers (webhooks, terminal runs): heartbeats are time-based polling, webhooks are push. Paperclip’s “Routines (beta)” feature supports both [src-001].
Related entities
- OpenClaw — originator of the heartbeat pattern
- Paperclip — current implementation
- Claude Code — the runtime each heartbeat wakes up into
Related concepts
Source references
- [src-001] Nate Herk — “Claude Code + Paperclip Just Destroyed OpenClaw” (2026-03-28)