A heartbeat can make OpenClaw feel autonomous, but it can also create silent runaway loops. This playbook shows how to design a heartbeat that is useful, cheap, and safe to run unattended.
Start with a heartbeat that only reports status. Then add one small action. Then add approvals for risky actions. Autonomy is earned.
The number one heartbeat mistake is skipping budgets and stop conditions.
OpenClaw supports both heartbeats (a periodic tick) and cron jobs (scheduled runs). A heartbeat is best for “check and report” loops. Cron jobs are best when you know the schedule and want explicit execution windows.
If you want to run a workflow once a day at a specific time, use cron. If you want a lightweight health check every N minutes, use heartbeat.
A good heartbeat is a bounded program. It should have hard limits and clear “done” states.
Budgets prevent silent spending and reduce the chance of retry storms when tools are flaky.
Design principle
A heartbeat should never be the first time you try a new tool. Prototype manually, then automate the stable version.
Use this template to avoid the classic failure: the heartbeat does “a little bit of everything” and becomes impossible to predict.
The key is to separate: observation, planning, and acting.
Heartbeat run format (output contract)
## Status
- What changed since last run:
- What looks risky or blocked:
## Proposed actions (no side effects yet)
1) ...
## Needs approval
- ...When tools fail, retries are necessary. But retries without limits create storms that look like “OpenClaw went crazy.”
Treat retries like you treat budgets: explicit, limited, and visible.
Parallel sub-agents can speed up research and triage, but they also multiply cost and failure surface area. Use them when the work splits cleanly into independent chunks with bounded outputs.
If you cannot define the chunk output in a paragraph, do not parallelize it.
Automation amplifies infrastructure problems. If your gateway is slow or frequently interrupted, heartbeats and cron jobs will look unreliable even if your workflow design is solid.
Clawdguy gives you a stable OpenClaw runtime so your automation failures are actually workflow failures, not infrastructure noise.