Magatama, From the Beginning, Part 2: The Chore I Kept Choosing

Part one of this series covered a diagnosis-only agent, kept narrow on purpose because I didn't trust myself to get the guardrails right on the first try.

That version held up fine. For a while I assumed I'd leave it exactly where it was: read the output, then handle it myself. No execution, no exceptions.

What changed my mind wasn't a failure. It was a pattern in my own behavior. Magatama would flag something, I'd read the diagnosis, agree with it completely, and then go type out the same handful of commands I'd typed the last time the same category of problem showed up. Same fix, same sequence, same hands. After enough repeats of that, the no-execution rule stopped feeling like caution and started feeling like a chore I kept volunteering for on purpose.

So I went back and looked at the scope question again.

Two options I turned down

There were two obvious ways to widen it. One was full autonomy: let the agent diagnose and act on anything it judged safe, with no human anywhere in the loop. The other was a whitelist: sit down ahead of time, pre-approve a set of actions, and let it execute the moment something matched, again without asking anyone.

I turned down both.

Full autonomy means trusting a diagnosis I haven't even seen yet with an action I can't take back. Handing over that much trust in one step doesn't get easier to justify just because the agent has been right before. Being right in the past says nothing about the case that matters most.

A blind whitelist looked safer on paper, but it just moves the risk earlier, to the day I sit down and write the list. On that day I'm reasoning about categories of problems in the abstract, not the specific one sitting in front of me with its own logs and its own context. I'd rather make the judgment call with real evidence in hand than have already made it weeks earlier on a hunch about what "safe" tends to look like in general.

Diagnosis, plus a named playbook

What I landed on sits between the two. The agent still diagnoses first, same as before, no change there. If it also has a fix in mind, it proposes that fix alongside a specific, named playbook for carrying it out: an actual documented procedure written in advance, spelling out exact steps, rather than a vague suggestion to restart something and hope.

Then it waits. Either for me directly, or for a narrow, pre-approved rule that covers the low-risk cases where I've already decided a given playbook is safe to run unattended. Everything that executes gets logged, step by step, whether I approved it in the moment or it ran under a standing rule I set up in advance.

The difference between this and the whitelist I turned down is where the judgment sits. A whitelist asks me to pre-clear actions in the abstract, once, and then trust that abstraction forever. This setup asks me to pre-clear procedures instead, and still requires either my sign-off or a narrowly scoped rule before any one of them runs against a live diagnosis. Each playbook exists because I sat down and thought through one specific failure mode on its own terms, not because I was trying to cover every possible case in advance.

Magatama stopped being purely a diagnosis tool right there. It became something that could change things on my own infrastructure, inside limits I'd written down ahead of time and could still go back and revise whenever they turned out to be wrong.

Part three covers the week that put this design to the test for real.