OpenClaw subagents, tool policies, and the file 90% of people get wrong

Your agent doesn't know what it's NOT allowed to do...

8 days in and your agent is actually doing things on its own.

But here's the problem - it's unsupervised. No priorities, no delegation, no guardrails.

Day 9 is where that changes.

Most people think the soul.md file is where all the important config lives. It's not. agents.md is what your agent reads first, every session, including every background task it spawns. Get this file wrong and everything downstream inherits that mess.

The part I find most underrated is subagents. You can literally have your agent spin off a worker mid-conversation, keep going without waiting, and get a summary dropped back in when the job's done.

And the security model is cleaner than most people expect. Three independent layers: where code runs, which tools are available, and who can override. They don't overlap, they stack.

Day 10 we're getting into tools and skills, so hat your agent can actually do beyond the defaults.

🎥 Watch Day 9 below.

3 comments