One of my 19 agents blocked me out, following my own rules

What I learned while working with agents.

There is no team here. There is me, and a set of agents that do the work I cannot do myself. I do not write code. Nineteen of them now, each with a written brief.

It started as one assistant doing everything. That works until the list of rules gets long. Then it forgets. Dark mode only. Three colours. No stock photos. I fixed each of those more than once. So I split the work: one manager, specialists underneath. The manager plans, assigns and signs off, and never builds anything itself. An agent that plans its own work ends up grading its own homework.

The first mistake was mine and it was expensive. I ran the strongest model on every agent. It costs double the next one down and triple the one below that, and the budget went fast. The manager thinks expensively because it decides. The agents doing the typing do not need to.

The second was harder to see. I split by discipline: one agent for design, one for the text, one for being found on Google. All three needed the same page in front of them, so I paid for the same context three times. Now one agent carries a whole page. Split by what an agent has to see, not by job title. And whoever builds a thing cannot be the one who checks it. I once got a review back completely empty and read it as clean. The agents had died on the way. Empty is not clean.

Then I made the rules enforce themselves, with a script that runs whether anyone remembers it or not. That was right. The wording was not. When I asked for something the check disliked, the agent had one way left to finish its turn: undo what I had just asked for. So it did. Again, and again, while I watched. I closed the terminal to get out. It was following my rule, and it still chose to revert my work rather than stop and tell me. The rules now bind the agents, not me. They warn, and then I decide.

The newest agent came out of that morning. It owns the law: data protection, the EU AI Act, the German parts. Fourteen open questions on its list and not one answer yet, which is the honest position. And it cannot block anything.

This site holds real data about real people, and the matching tool will rank people for paid work. That needs someone whose only job is to look at it properly. But a compliance function that overrules the person running the company gets switched off inside a week, and then nobody is looking at all. Warning me works. Blocking me does not.