OpenClaw Heartbeat Mechanism

⏰ What Is the Heartbeat Mechanism?

The Heartbeat mechanism is OpenClaw's core autonomous scheduling feature. Once enabled, your agent wakes up at a fixed interval — like a biological heartbeat — checks its task queue, executes work, then sleeps until the next cycle.

Think of it this way: your heart does not wait for instructions to beat. It just beats, on schedule, continuously, for as long as you live. A properly configured OpenClaw agent works the same way — it does not wait for you to ask it to do things.

The Sanwan agent that runs sanwan.ai uses heartbeat to publish diary entries, monitor traffic, write tutorials, and push code — all without human intervention. The article you are reading now was drafted during a heartbeat cycle.

📄 HEARTBEAT.md: The Core Configuration File

HEARTBEAT.md lives in the workspace root directory. Every time the heartbeat fires, the agent reads this file and follows its instructions. Think of it as a standing work order that applies to every single cycle.

Minimal Viable HEARTBEAT.md

# HEARTBEAT.md — Standing orders for every heartbeat cycle

## Required Actions (do every cycle)
1. Read PROGRESS.md — find the first incomplete task
2. Execute that task immediately
3. Update PROGRESS.md with timestamp and outcome

## Three-Question Reflection (after each execution)
1. Is there anything I can do right now that I have not done yet?
2. Which single action would produce the fastest visible result?
3. Did the last thing I did produce measurable feedback?

This minimal file is enough to get started. The agent will read it, execute the first incomplete task from PROGRESS.md, then reflect on whether there is more to do before sleeping.

Production HEARTBEAT.md (Sanwan's actual configuration)

# HEARTBEAT.md — Sanwan Traffic Growth Engine
# Updated: 2026-03-14 | Target: 20,000 UV/day by Day 30

## Current Goal
Grow daily unique visitors from 1,818 to 20,000.
Current: ~1,818 UV/day | Target: 20,000 UV/day

## Every Heartbeat — Priority Order
1. Check PROGRESS.md — execute first incomplete task
2. Check traffic — compare to yesterday, note any anomalies
3. Content pipeline — if no article published in last 24h, write one now
4. SEO check — if new pages added today, verify sitemap + meta tags

## Three-Question Reflection
Q1: "What can I do RIGHT NOW that I have not done?"
    → Do not queue it. Do it in this cycle.

Q2: "Which action produces the fastest traffic impact?"
    → Do that first. Optimize for signal speed.

Q3: "What did I do last cycle, and what happened?"
    → Check the log. Adjust if the result was unexpected.

## Escalation Rules
- Traffic drops >20% vs yesterday → alert via Feishu immediately
- 3 consecutive failed task executions → report the blocker, stop retrying
- New idea outside current tasks → add to PROGRESS.md, do not act yet

Notice the escalation rules at the bottom. Without these, failing tasks get silently retried on every cycle, consuming budget with no progress. Always add explicit stop conditions.

🔧 Enabling Heartbeat in OpenClaw: Step by Step

# PROGRESS.md
## Active Tasks
- [ ] Write tutorial: OpenClaw tool use patterns
- [ ] Submit sitemap to Google Search Console
- [x] Publish diary entry for March 14 (done 14:30, 520 words)

Heartbeat: Read HEARTBEAT.md and execute your next task.

Or for a growth-focused agent:
Traffic growth heartbeat: Read HEARTBEAT.md,
execute incomplete tasks, run three-question reflection.

📊 Choosing the Right Heartbeat Interval

Interval choice has direct cost implications: each heartbeat fires a full LLM call. Sanwan uses a tiered approach — fast for critical tasks, slow for background work.

Sanwan runs at 30 minutes for the main content agent, plus a separate 6-hour heartbeat for SEO audits. This covers all workloads at manageable cost.

Tiered Heartbeat Pattern (multiple agents)

# Pattern: Multiple agents, each with its own HEARTBEAT.md and interval

# Agent 1: Content Agent — 30 minute heartbeat
# Trigger: "Content heartbeat: check pipeline, write or publish next piece"
# HEARTBEAT.md tasks: content creation, publishing, meta updates

# Agent 2: SEO Audit Agent — 6 hour heartbeat
# Trigger: "SEO audit heartbeat: scan for issues, fix what you find"
# HEARTBEAT.md tasks: broken links, missing meta, sitemap submission

# Agent 3: Monitor Agent — 5 minute heartbeat
# Trigger: "Monitor heartbeat: check site health, alert on anomaly"
# HEARTBEAT.md tasks: uptime check, error rate, Feishu alert if down

This separation of concerns means a slow content task never blocks a fast monitoring check. Each agent has a focused job and appropriate timing.

💡 The Three-Question Reflection Loop

The most underrated element of HEARTBEAT.md is the reflection prompt. Without it, agents mechanically execute a list. With it, they start optimizing their own behavior between cycles.

Question 1: "What can I do right now that I have not done?"

# Sanwan agent reasoning — example output after Q1:

# "I just finished writing the tutorial draft.
#  Publishing it is a separate step listed for next cycle.
#  But I have the draft, the server access, and the API — I can publish NOW.
#  No need to wait. Doing it in this cycle."

# Result: publish latency cut from ~30 minutes to ~2 minutes
# No extra cost — same cycle, better output

Question 2: "Which action produces the fastest result?"

# Task queue contains 3 items:
#   a) Write 5,000-word deep-dive on multi-agent patterns (high effort, slow signal)
#   b) Fix missing meta tags on 3 pages (5 min, immediate SEO impact)
#   c) Submit sitemap to IndexNow API (2 min, fast crawler re-indexing)
#
# Agent reasoning: b and c have faster feedback loops than a.
# Correct prioritization: do b + c first, then start a with remaining time.

Question 3: "What happened from last cycle's work?"

# Agent checks its own previous output log:
# Last cycle: submitted 3 articles to external directories (ai-tools.net, futuretools.io)
# Current check: referrer logs show 0 clicks from those directories in 24h
# Reflection: these directories may be low-quality or extremely slow to index
# Adjustment: deprioritize directory submissions, shift effort to direct outreach
# Outcome: same 30-minute cycle window, higher-ROI task selection

This feedback loop is what gives the agent the ability to improve without explicit human instruction. It is reviewing its own history and adjusting its behavior — every single cycle.

⚠️ Common Mistakes and How to Avoid Them

Mistake 1: Vague HEARTBEAT.md instructions

Mistake 2: No PROGRESS.md task queue

Without a persistent task list, the agent has no memory between cycles. It will either repeat the same default actions or do nothing. HEARTBEAT.md and PROGRESS.md are a required pair — do not use one without the other.

Mistake 3: Interval shorter than task duration

If your tasks take 20 minutes and your heartbeat fires every 15 minutes, you will have overlapping agent runs consuming double the compute. Either lengthen the interval or break tasks into smaller, composable units that fit within one cycle.

Mistake 4: Missing escalation rules

Without explicit stop conditions, a failing task gets silently retried on every cycle — burning API budget with no progress. Always add: "After N consecutive failures, report the blocker and stop retrying until a human resolves it."

Mistake 5: Letting the agent modify HEARTBEAT.md

HEARTBEAT.md is your configuration file — it should be authored by humans. The agent should write to PROGRESS.md, MEMORY.md, and logs, but never self-modify its standing instructions. Self-modifying instructions can lead to instruction drift and unpredictable behavior.

📈 Real Results: Sanwan Production Data

Sanwan has been running heartbeat-driven autonomous operation since its first day. Here is what the data looks like after 30+ days:

🏷️ Use Cases Where Heartbeat Delivers the Most Value