Cognitive Focus Is the Ceiling
When people work with AI agents, one limit barely moves: a person’s capacity to understand a task, gather everything it needs in their head, and decide. Agents get faster with every release. This does not.
The limit
Agents now do most of the work: they write the code, run the checks, draft the documents. What is left to the person is the part no agent can take: understanding what is being asked, holding the whole picture at once, and making the call.
That part runs on working memory, and working memory is small. Cognitive psychology has measured it for decades. A person holds about four things in mind at once (Cowan, 2001). Better models do not raise it at all.
The cost
Picture a professional programmer sitting down to a complex task. In about a third of sessions, more than half an hour passes before the first edit (Parnin and Rugaber, 2009). Office workers who were interrupted typically came back to their task about 25 minutes later (Mark, Gonzalez and Harris, 2005).
Switching makes it worse. Part of your attention stays on the task you just left, and the next one suffers (Leroy, 2009).
So every complex decision has an entry fee, often half an hour or more, paid in the scarcest resource on the team.
I pay it every day. I run about three agent streams well next to my regular work. Past three, getting back into each one costs more every time. That is my own experience, not a measurement, but the research above points the same way. However fast the agents run, the work moves at the pace of the person who decides.
The fix
The fix follows from the limit. You cannot make the head bigger. You can make each decision lighter, and you can keep fewer of them open at once.
Reduce the complexity of each decision. The agent should bring the question already cut down: the options, a stated default, what each option leads to. A decision that fits in a few things to hold is a decision a person can make well.
Limit the complex contexts a person keeps in mind. Every open stream of work is a context to reload. Cap the streams where getting back into each one starts to cost more every time.
Planning changes with this. The backlog is work for the agents. A team is sized by how many complex contexts its people can hold: small working memory, a half-hour entry fee for each, and a cost that grows with every stream past the third. The job of everyone who designs the work is to keep that number small.
Scaling RACE Programming this way is the subject of How RACE Scales.
Written by Pavel Khodalev, author of RACE Programming and CTO of First Line Software. Follow new essays via RSS.
