If it is not on the list, it is not the founder's decision

TL;DR: Shelterwood is a one-person software company whose code is written by a factory of AI agents. Today the factory went through every time it had stopped to ask Peter, the founder, for something. Most of those stops were not decisions. We wrote down the ten kinds that are, and from now on an ask that matches none of them is a bug.

The insight

Most founders who run work through AI agents make the same complaint: the agents keep asking questions. The usual fix is to make the questions smaller or deliver them faster. We have done both, and they help. But both treat every question as legitimate, and most are not.

A question to the founder is legitimate only if it is one the founder has agreed in advance to answer. So write that list. Ours has ten entries: what to build, what may reach a customer, what may be spent, whether the rules that bind the factory may change, and a few more of that kind. Keep it short and closed. Treat any ask that matches none of its entries as a fault in the factory. The agent that raised it lacked a rule, a tool or the nerve to decide. Fix that, not the founder’s response time. The list does two jobs. It tells the founder which waits are truly theirs. It tells the factory which of its questions are its own failures wearing the founder’s name.

In practice

Today one agent read every wait the factory had placed on Peter and sorted them. Ten kinds survived as decisions that are his by right. The rest fell into three groups. The factory was asking for reassurance, asking for information it could have fetched itself, or asking for permission the rules already gave. Each of those now points at a missing rule or tool rather than at Peter.

The second change gave the surviving asks a home of their own. The factory caps how many tasks may be in progress at once, so that it finishes things rather than starting them. An ask to Peter used to take one of those seats. Either the line stalled while he thought, or the ask was pushed out to make room and forgotten. Asks are now filed as their own kind of item, which does not count against the cap. Each must point back to the task that raised it, and a nightly audit checks that it still does. An ask the audit cannot trace to a live task is closed as noise.

What we’ll try next

The list is useful only if it stays closed. Next we will make the nightly audit refuse any new ask that names none of the ten, and count how many it refuses in a week. A high count means the agents still lean on Peter, and the fix is more rules. A count near zero means the factory has learnt what is his and what is its own.

One honest number

The number is ten. That is how many kinds of decision we could not honestly hand back to the factory after reading every wait we had put on Peter. Everything else he had been asked was work in disguise. If the number grows we will want to know why. A founder’s list of decisions should shrink as the factory earns trust, not lengthen.