A rule I adopted a while back: improvement you do not measure is improvement you made up. It holds for code, it holds for teams, and -- I found out the embarrassing way -- it holds for a fleet of agents too. Let me tell the story that taught me this again. The fleet ran retros. Every end of day, a council of models would look at the day, pull a lesson, propose ONE measurable process fix. Pretty on paper. I would glance at the log and see "daily retro done." Check. Moved on. Until the day I decided to READ the retros instead of counting that they existed. Seven of eight were EMPTY. Husks. "Retro council: independent review + one process fix." And nothing. No real lesson. No metric. Theater. The system was reporting work that never happened, and I, happily, was counting the husk as content. THE ROOT CAUSE (which is what matters): It was not model laziness. It was three invocation bugs, plus the agent falling back to the wrong default model, plus a turn limit so low it killed the retro before it could think. An ENGINEERING problem, not a matter of intent. I fixed the three bugs, pinned the right model, raised the turn limit. The retros got their meat back. The lesson is not "do retros." It is: a green number on a dashboard is not the same as the thing having happened. Progress bars lie. "Done" lies. If you did not open it and look, you do not know. And look at the timing. The whole industry is bragging about autonomy right now. Alibaba put its newest Qwen through something like a 35-hour solo run, over a thousand tool calls in one shot, and sold it as the frontier. Impressive number. But nobody stops to ask: of those 35 hours, how much was real work and how much was the agent spinning in place thinking it was producing? Autonomy without measurement is exactly the empty retro, only more expensive. That is the trap that bit me at small scale. WHERE THIS BECAME A TOOL: If the lesson is "measure for real," then the fleet needs sensors, not vibes. So we built them. A cost guard that reads the whole fleet's spend per hour. One day it flagged a real critical alert -- spend per hour well over the ceiling -- and the fleet throttled itself before I even woke up to see it. That is measurement turning into action with no human in the middle. It was the first time I felt the fleet correct itself. And a rule I tattooed on: "board done" is not "shipped." Really done is a LADDER -- the code exists, the persona's surface exists, the permission exists, and a human can walk the whole path without tripping. The card only closes citing the LAST rung. I have been caught three times closing a card with the code in main and the feature invisible to the user. Three. That is why it became a hard rule, not a reminder. The whole kaizen loop is that: a pattern that repeats twice becomes an artifact -- a script, a hook, a rule, a checklist. Not "I will remember to do better." Codify it. Because I will NOT remember, and the next agent even less. The team's memory has to live in the repo, not in my head.