The heartbeat pattern was named in 2024 and became mainstream in 2025 through BCG's 'New Wave of Autonomous AI Agents' brief and
Paperclip's open-source documentation. Before heartbeats, AI agents could only respond to human prompts — which meant they couldn't run a business, because a business runs on schedules (daily briefings, weekly reviews, monthly close, quarterly planning).
A heartbeat has four parts:
1. A trigger (cron schedule, event listener, or external webhook).
2. A context refresh (pull current state from Notion, Postgres, Stripe, etc.).
3. A task execution (the agent does the work, writes outputs back, possibly escalates).
4. An audit trail (what ran, what was produced, what got flagged).
Practical examples in a
one-person company:
- Daily 7am:
AI CEO reads overnight metrics, drafts the founder brief.
- Weekly Monday 9am:
AI CMO reviews last week's content performance, proposes this week's calendar.
- Monthly first-of-the-month:
AI CFO closes books, generates P&L, flags anomalies.
- Event-triggered: new customer signup → AI Support agent fires, sends onboarding sequence.
Heartbeats are what separate 'AI tool' from 'AI employee.' A tool waits. An employee works.
Implementation varies:
- Paperclip has explicit heartbeat configuration per agent.
- Polsia runs continuous autopilot (everything is a heartbeat).
- Tycoon exposes heartbeats as scheduled tasks per role that can be paused, tuned, or audited.
Best practice: every AI role should have at least one heartbeat. Roles without heartbeats never volunteer work and effectively operate as chatbots. The transition from 'AI helper' to 'AI teammate' is when heartbeats land.