It’s not enough to write a simple prompt to make your @openclaw agent work while you sleep. The only way to make it run for a long time autonomously is to build a proper loop infrastructure. A single prompt works like this: you say "build X", the agent works for a while, hits a bug it can't solve, and stops. So instead of one long session, you run the agent in short cycles with a file it reads and writes to between each run. 1. Create a task file tasks(dot)md that describes what needs to be done, step by step 2. Create a rules file context(dot)md with your stack, constraints, and known pitfalls, things like "always restart pm2 with --update-env" or "db url must be localhost not docker host" 3. Run the agent on a timer. Every 10-15 minutes, a cron job triggers it. each cycle: > reads tasks(dot)md → picks the next incomplete step > reads context(dot)md → knows the rules and past mistakes > works on that one step > writes the result back: "step 3 done" or "step 3 blocked: port 3000 already in use" > if blocked, writes WHY it's blocked and moves to the next step Obviously you can use the help of your agent for all of these. 4. The key is when it fails, it doesn't stop. It logs the error, skips to the next task, and tries the next cycle again. The failure becomes the context for the next run. Single prompt agent is fragile, one error kills the whole run. A loop agent is resilient, errors get logged, skipped, and retried with a new context. After 8 hours of 15-minute cycles, that's 32 runs. Each one reads what the last one did, learns from its failures, and pushes forward.