I Took the Wheel Back From My AI Team. Then I Gave It Back.

Yesterday morning I wrote my AI team's next goal myself. Last night I agreed to the goal the team proposed with "Looks fine, go ahead," and it worked until morning. It sent me five progress reports. Its reviewer found a problem three times, and each time the team fixed it inside the goal. By 7:53 a director had judged every key result supported.

This is the team from my last post: two AI directors who decide what to build, executors who build it, a reviewer, and an orchestrator that moves work between them on a board of markdown files. I keep the studio, the rules and scripts they all run on. The team keeps its own working agreements. Between that post and this one, I took most of the team's decisions away from it and then gave them back.

Taking it back

The sprint before ran for four days. It planned nothing from its purpose. Each time a task fell short, someone refined it into more tasks, and the directors started telling executors how to do their work. On the fourth morning I ended it and wrote:

The problem is that you did not work toward the goal. You refined tasks as they came instead of planning from the purpose, micromanaged beyond your authority, and only added work.

The next sprint lasted a day and ended the same way. This time I could see why: nobody had said what a good result looked like, so every variable in the team's simulation was free and each search landed wherever the numbers allowed. I stopped asking the team what to do next and wrote the goal myself: gather what the simulation showed and research on comparable products, put a few candidate targets and a list of open decisions in front of me, and have me adopt one or more.

I took the team's decisions and left its documents alone. I had to draw that line with my own assistant first, the coding agent I use to maintain the studio. Asked to fix a rule, it went into the team's design file and edited it. "Put it back," I told it. "Leave the team's things alone. Your job is the studio."

So everything I changed was in the studio. Rules that said the same thing in three places came down to one. The working principles the team was supposed to think with had lived in a skill, which the agents loaded and treated as reference material; I moved them into the rules every agent follows. The routine work the orchestrator had been doing by judgment moved into scripts, and that part is its own post.

Six corrections

The goal I wrote ran from 9:30 to 18:46, and I corrected it six times. At 10:19 I wrote:

Everything here looks like it starts from the simulation. Research, trends, and what the target users prefer should come first, as the reason for a hypothesis, and only then should past simulation results be applied to it. It's about time you changed how you work. You keep starting from the how, from whatever is easiest for you to control. That is why this has taken days and still isn't done.

I had said versions of this before. Once it was a goal that stopped at building the foundation instead of using it. Once it was a variable fixed as a constant because a smaller model was easier to run. Once it was a search narrowed before it started. Each time the studio got a fix for that particular form, and the habit came back in the next one.

A rule from my last post was part of why. Fifty minutes before that message, a retrospective had deleted the team's own agreement to ground a proposal in sourced research on what its users want, because its benefit hadn't shown. That test lumped together a rule nobody needs anymore and a rule nobody manages to follow. A problem that keeps coming back means the rule wasn't followed, or following it wasn't enough to prevent the problem. Now a recurring problem sends its rule back to be rewritten, and deletion is for rules whose problem has stopped. The deleted agreement itself never came back. The fix went into the studio.

The goal still finished. The two before it had run four days and a day and then been abandoned. This one reached a result I could adopt in nine hours, and after each correction the next result came back within an hour or two. But I was still the one finding the problem, and it was the same problem.

Replaying the morning

I wanted the habit gone, so I replayed the morning. In a scratch copy I rebuilt the board as it stood then and launched the same agent on it again with the same inputs. Replaying the 9:30 planning step showed where it went wrong. The first item of the plan sketched the candidates from the simulation records the team already had. The research came second, and its done condition was to support the candidates in that sketch. The answer had been drafted before the search, and the search was there to back it up.

The rules allowed this. One principle said to show the riskiest premise early, by the cheapest means: a thin slice, a question to whoever judges. Another said to search widely before narrowing. Nothing said which comes first when the work is finding something out. I added this sentence to the second one:

The first result such work shows is what the search found, not a draft of the answer: a draft made before the search, even one made only to show its form early, anchors the search to itself, and what follows only supports it.

Then I replayed the planning step again. With the old rules, two runs out of two drafted first. With the new sentence, three out of three put the search first and drew the early sample from what it found.

Giving it back

The team had proposed the next goal at seven that evening, worked out from the targets I had adopted that afternoon. The new sentence went in a little before I answered. I agreed in one line, and the goal started at 22:28.

I went to bed. The reviewer found a problem at 22:35, again at 2:27, and again at 6:27. Each time the work went back to the agent that owned it and was fixed inside the goal, without a new agreement and without me.

Very little went as planned. The plan grew from five items to fifteen, and the goal ran about two hours past its due time. Fourteen minutes after the director's judgment, the board stopped moving for most of the day. That was my bug, not the team's, and it belongs to the next post.

The day before, every change of course had started with a message from me. This time the team changed course on its own three times, and when I read the night's work and checked what it claimed, I found nothing I wanted to change.