On a Sunday morning I spent thirty minutes making decisions and then walked away for four and a half hours. When I came back, a failing 1.52 TB drive with 118,599 files had been emptied and every file verified. About 8.43 TB of single-copy exposure was mapped, and a rescan I would have done by hand had been avoided by about 97 percent, simply by comparing names, sizes and dates first.
That was not luck, and it was not the AI being brilliant. It was rules I wrote down after the day an agent started five scanners without asking me. I rewrote the brief that night and added a line I have not had to repeat: never start a long job without approval.
The ladder
Everything an agent might do sits on a five-step ladder. The higher the step, the more a person has to be in the loop.
- Look at things: AI can do it anytime.
- Change reversible tooling: AI, with a written grant.
- Make runtime-only changes: by explicit, scoped permission.
- Touch data: a person pastes a command the AI prepared, and it ends in a clear GREEN or RED.
- Destroy something: a human only, after identity checks and with verified copies elsewhere.
The habits that make it work
- Every approval is pasted word for word into a log. There are 113 of them so far.
- Every status line points at a file or a count.
- A denial is a hard stop. It is never an invitation to find another route.
- There is a STOP word, and a launcher that refuses to start a job that is already running. It stopped me once, when I pasted the same command twice.
I like AI that works while I am away. I like it a lot more when I know exactly what it is allowed to do. If you are thinking about putting agents in front of your own data, start with the ladder and not the model. The details are on the guardrails page.