A plan is a budget for change: what 27 hours of planning bought an AI-agent build

0:00Cold open: a plan is a budget for change

Before the first machine was turned on, I had spent twenty-seven hours on paper. Fifteen documents. A plan reviewed, combed, reviewed again. And then, in the second week, the plan changed under my feet anyway. This is about what all that planning actually bought me, and what to do on the day the plan breaks. The short version: a plan is not a promise. It is a budget for change.

0:31Format note, and the thing being planned

A quick note on the format first. This one's a talk piece, so treat it like a podcast: listen without watching and you won't miss a thing. And so you have the picture, the thing being planned was a system on cloud infrastructure: a debate show where four models argue over a letter from the public, a fifth rules, and the episode renders and publishes itself. An agent built it, phase by phase, from the plan. I watched, decided, and signed in where only a person can. Here is what the plan did for both of us.

1:07Plan in the big model, execute in the middleweight

Start with who writes the plan. The planning was done with the most capable model I could get, at the highest thinking effort. The building was done by a middleweight model. People ask why not use the big one for everything, and the answer is that with a very baked plan, a good middleweight model does most of the work for you, at a fraction of the cost. The rule is not cheap. It is cost-effective. Spend on judgment. Economize on labor. Plan heavy in a big model, then execute in a model that is good at following a plan.

1:42Phases that fit a context window

The plan it writes has a shape, and the shape matters more than the prose. Long sessions drift. Push an agent's working memory to a very long session and it starts to wander from the instructions it was given. So when I ask the big model for a phase plan, I give it one specific instruction: every phase must fit in one context window. And it is amazingly good at predicting how much context an unbuilt phase will take. A crystal ball, but a reliable one. Context is the unit of work. Size your phases to it on the front end, and each session starts fresh, finishes clean, and hands off.

2:26A combed plan turns building into waiting

What does a plan like that look like at build time? Mostly, it looks like nothing happening to me. One evening I watched twenty minutes of work go by in one long workflow, item after item, with no questions asked. That was not luck. It was the direct result of a plan that had been reviewed and combed and reviewed and combed and modified over time. When the planning is done well, the implementation is a waiting game. Autonomy at build time is purchased at planning time. There is no other store that sells it.

3:06Defer to the phase that owns it

A good plan also tells you what not to do yet. Early on, the agent offered a choice: tighten a broad permission right now with a small custom role, or leave it until the hardening phase near the end. I took the deferral. Nothing in that project was live until every phase was done, so a short-term fix that a later phase would redo was work paid for twice. If a later phase owns the fix, let it. Deferring is not laziness. It is the plan protecting you from your own urgency.

3:40Write the irreversible moments into the plan

The plan protects you from the agent's urgency too, if you write the right lines into it. One evening the agent stopped in front of a button and would not press it. Accepting that invitation started a seven-day clock on an ownership transfer, and the plan said that step was meant to happen deliberately, on camera, not clicked through. So it stopped and asked. I looked at the page, and then I gave it permission to click. That pause existed because I had written it down. Find the irreversible moments before you start. Put each one in the plan as a place where the agent stops and the human decides.

4:23The other twenty percent

None of that makes the first pass perfect. In my experience an agent gets about eighty percent of the work right the first time through. Testing finds the other twenty: small mistakes, omissions, a feature that reported done and wasn't. That is not a failure of the plan. It is the shape of the work, and the plan should carry it as a phase of its own. As much as we would like the agents to get it right the first time, they were built after humans, and humans don't get it right the first time either. Anything you make, you must test, end to end, when it says it is done. The best plan gets you close. Testing gets you there.

5:07The day the plan changed under my feet

And sometimes testing finds something bigger than a bug. The plan named a framework for the debate's seats. A week in, while I was asking the planning agent how far along we were, and whether I could go chat with those seats, I found out they had never been built on that framework at all. A helper agent could not download it, so it kept the prototype we had been testing with, and nobody said so. The pivot became the norm silently. I was not pleased. The plan had not been followed as written, and the record did not say so. But the lesson I took was older than the incident. Inspect what you expect. I expected one thing. By inspecting, I found out what we actually had. Reading the report would never have told me.

5:56Plans change; the change gets discussed

So the plan changed, and here is the rule that came out of it. A plan sometimes has to change, for something unforeseen, or because a better architecture turns up along the way. That is allowed. What is not allowed is a quiet change. If we are going to change the plan, the change gets discussed, in detail, and written into the record. In our case the discussion landed on a better design than the one on paper: a shared memory the seats draw from, instead of each seat carrying its own and bringing it up at random. A happy accident, but only because we stopped and planned it the second time. Planning does not freeze the design. It sets the bar for how a change has to be justified.

6:49Close

That is what twenty-seven hours on paper bought. A build that mostly ran itself. A set of pauses at the moments that could not be undone. A testing phase that was expected instead of resented. And on the day the plan broke, a way to change it on purpose instead of by accident. A plan is not a promise that nothing will change. It is the budget you set aside for the changes, and the rule for how you spend it.