How do you make OpenClaw manage long-running tasks proactively instead of acting like a one-shot cha

[fresh ibex]

Hi everyone,

I’m trying to make my OpenClaw agent behave more like a proactive long-running agent, not just a chatbot that replies once and then waits for the next instruction.

Right now, when I give it a larger task, such as building a demo site, installing a scraping tool, investigating an issue, or tracking something over time, it usually only responds in the current chat or executes once. It does not automatically create a persistent task/project folder, save requirements, track progress, update any task list, or schedule follow-ups with Heartbeat/Cron.

I’m trying to understand the correct workflow for this.

My questions are:

1. How should I structure long-running tasks in OpenClaw?
2. Should I create something like TASKS.md plus per-task folders with TASK.md, PLAN.md, PROGRESS.md, AUTOMATION.md, and HANDOFF.md?
3. Should I define a “Task Intake Protocol” in AGENTS.md, so when I type something like Task Create: [description], the agent automatically creates the task structure and tracks it?
4. How should Heartbeat and Cron Jobs fit into this?
5. Should Heartbeat read a central TASKS.md and check tasks marked heartbeat_watch=true?
6. When should I use Heartbeat vs isolated Cron vs custom-session Cron vs Task Flow?
7. Is there any built-in or recommended way for OpenClaw to detect “large tasks” during chat and turn them into persistent projects automatically?

My goal is to have the agent keep track of unfinished work, continue from saved progress, and proactively check or resume tasks when appropriate.

If anyone has a working setup, templates for AGENTS.md / HEARTBEAT.md / TASKS.md, or a recommended folder structure for long-running tasks, I’d really appreciate seeing how you’re doing it.

[junior adder]

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

[undone hamlet]

Short version, you’re not missing some hidden “long-running agent mode”. OpenClaw gives you the primitives, but not an opinionated built-in project manager.

What’s built in:

• AGENTS.md for standing orders / operating rules
• HEARTBEAT.md for periodic checks
• cron for exact scheduled work
• background tasks as the runtime ledger
• Task Flow for durable multi-step orchestration

What is not built in:

• automatic “this looks big, create a project folder + plan + follow-ups” conversion
• automatic loading of arbitrary files like TASKS.md or tasks/foo/PLAN.md

Important bit: only bootstrap files are auto-injected every session, things like AGENTS.md, HEARTBEAT.md, MEMORY.md, etc. A custom TASKS.md or tasks/... folder is fine, but you need to teach the agent to use it from AGENTS.md.

My recommendation:

1. Structure long-running tasks with two layers

• Runtime layer: use cron, heartbeat, tasks, Task Flow
• Human/project layer: keep your own lightweight task files

2. Yes, use a task folder pattern, but keep it lean
   I would not start with 6 files for every task.

Better default:

• TASKS.md as index
• tasks/<slug>/TASK.md
• tasks/<slug>/PROGRESS.md
• tasks/<slug>/HANDOFF.md

Only add:

• PLAN.md for genuinely complex work
• AUTOMATION.md if the task has cron/heartbeat wiring

3. Yes, put a Task Intake Protocol in AGENTS.md
   That’s the right place for:

• trigger phrase like Task Create: ...
• when to create a task folder
• how to update TASKS.md
• when to add heartbeat vs cron
• when to escalate / ask for approval

4. Heartbeat vs cron

• Heartbeat: approximate, batched, context-aware checks
• Cron: exact timing, isolated or session-bound execution

5. Should heartbeat read central TASKS.md?
   Usually no. I wouldn’t make every heartbeat parse a giant task registry, that gets expensive and noisy.

Better:

• keep HEARTBEAT.md tiny

• use its built-in tasks: block for due-only periodic checks
• only mirror a few “watch” items there
• use cron for exact follow-ups

6. When to use each

• Heartbeat: “check inbox/calendar/status every so often”
• Main-session cron: reminders or “wake me next heartbeat with this”
• Isolated cron: reports, audits, standalone background jobs
• Custom-session cron: recurring work that benefits from persistent history
• Task Flow: multi-step pipelines where A leads to B leads to C
• Background tasks: tracking/audit, not scheduling

7. Built-in large-task detection?
   As far as current docs go, no. The recommended shape is to define conventions in AGENTS.md and enforce them with cron / heartbeat / standing orders.

A practical setup I’d use:

• AGENTS.md

  • define “if user says Task Create: or task is clearly multi-session, create/update task docs”
  • define intake checklist
  • define when to schedule follow-ups

• HEARTBEAT.md

  • tiny checklist or tasks: block only
  • not your full project database

• TASKS.md

  • one-line summary per active task
  • next action
  • status
  • owner / automation flag

• tasks/<slug>/...

  • details for active projects only