July, late, working on my own build.
For months I worked the way I think most people work. Open a session in the morning, describe the project, and keep going until the thing was finished or I was.
And every one of those long sessions had the same shape. The first hour was extraordinary. The middle was good. Then somewhere in the afternoon the quality fell off a cliff, and I would find myself repeating instructions I had given clearly two hours earlier.
For a long time I blamed myself. I assumed my instructions had drifted, or I had got sloppy as I got tired. That is a reasonable thing to assume and it was wrong.
What was actually happening is that the model has a fixed amount of room to hold things in. Everything counts against it. The rules I set at the start, every file I showed it, every reply it wrote, all of it accumulates in the same space. And as that space fills, the thing does not warn you. It gets lazy. It starts ignoring rules it followed perfectly an hour before, and it stays exactly as confident while it does it.
I went looking for a gauge. Something that would show me how full it was, a needle on a dial. There is no such gauge. You are flying it by feel.
So I learned to read it the way you read a person. Somebody who has been in a meeting too long starts agreeing with everything. They stop pushing back. Their answers get shorter and smoother and less useful. The model does the same, and once I knew to watch for it I could see it coming maybe fifteen minutes before it got bad.
Then I made two changes and both of them worked immediately.
The first was to stop treating the room as free. I started deciding in advance how it would be spent. A small slice at the top for the rules, the things the agent must never do. Then the largest share for the actual work. Then, and this is the part I had never done, a reserve I refused to touch, held back for the part where something breaks and you need room to investigate it.
That reserve is the whole trick. Before, I would fill the space with work and then hit a problem with nothing left to solve it in. The investigation is exactly when you need the most room, and I had been arriving there with the least.
The second change was smaller and it changed more. I stopped running one long session.
Now I finish a piece of work, and before I close, I have the agent write down where things stand. What was done, what was decided and why, what is still open. Then I close it and open a new one and hand it that note.
The first time I did it I expected to lose something. Continuity, context, the feeling that the thing knew me. What actually happened is that the new session was sharper than the old one had been for hours.
I described it to the class using a jewellery shop, because that is a room people can see. An associate has been with a customer for forty minutes. They know the budget, the anniversary, the ring size, which one the wife already rejected. Then their shift ends. They can walk out and let the next person start from nothing, or they can spend ninety seconds handing it over. Same shop, same customer, completely different afternoon.
The agent has a shift. It ends whether you plan for it or not. The only question is whether anybody wrote anything down before it did.
I keep a rules file now that the agent reads before it starts. It says what it must never touch. Production data. Anything that sends mail. Anything that deletes.
I wrote that file the week after the morning in Part Two.