Write It for Someone Who Remembers Nothing
Written later and filed to this date, as part of filling in every day of the year.
You asked what to set up first. Narrower answer: the one part I would build again before anything else, unchanged.
Every session your agent runs starts with no memory. Not a short memory: none. It will read whatever you put in front of it and believe that is the whole world. If the thing in front of it is the code, it is reading a terrible account. Code says what exists, never what was tried, never what was decided against, never what is waiting on you. A session that reads only code will rediscover Tuesday’s dead end with full confidence on Thursday.
So the one thing: give every repository a short file the next session reads before anything else, and make the agent update it last, every time, before it stops.
Mine has three sections and they have earned their names. Now: what is in flight, in two lines. Next: what to pick up, in order. Blocked: what is waiting on a human, and why, so that nothing with no memory tries to be helpful about it. That last section is the one that matters. A gate you forgot to write down is a gate that will be walked through politely at three in the morning.
Keep it short enough to be read in full. The moment it grows past a screen, sessions skim it, and a skimmed state file is worse than none because it feels read. Push the history somewhere else, a journal with a date on every entry, and let the state file point at it.
And the part nobody tells you: the file is for you as much as for the machine. Three weeks in, you will open a repository you have not thought about and you will not remember what was in flight either. The agent with no memory and the operator with a full life arrive at the same file with the same question, and it should answer both in under a minute.
I would not start with a fleet. I would not start with a dashboard. I would start with one repository, one agent, and a file called STATE that says what is true, updated last.