Call the Human Only to Decide
Agents should call a person only when a decision is theirs to make, and bring everything needed to make it in one reading. Everything else they do on their own, quietly.
Cognitive Focus Is the Ceiling named the limit: each person’s focus. This essay is about how to spend that focus on decisions and nothing else.
I mean the agents people already work with, in a chat: a coding agent in the terminal, an assistant in a desktop app. Most of what such an agent sends a person today needs nothing from them. It is a feed of output. The person scrolls it to find the one line that needs a decision.
I run a personal assistant built from agents. It is a first attempt, not a rule of the framework. The agents work on their own and call me only when a decision is mine. Everything else stays quiet.
The goal I gave the agent that helps me build it is simple. Keep me out of the routine. Bring me in only for what adds something new to the shared work, at the moment it is needed, through one channel, with as little load on my attention as possible.
The calls are small. Approving an outgoing message takes one reaction. A merge takes one word. A choice comes with a stated default, so I can accept it as it stands or pick another option.
A good call is answerable in one reading. It carries the question, the options, the default, and what each option leads to. When I have to act in another system, it brings a direct link to the exact screen. I answer with one word or one reaction, often from my phone.
I count it, week by week. The system is still in active development and tuning, so the numbers show my changes to it as much as the agents themselves.
Of the results the agents sent for my review, the share I accepted the first time, with no rework at all, went from 9% in the first week to 20% in the second and 25% in the third. Of all my answers, the share given with one reaction went from 0%, when that option did not exist yet, to about 25% and then about 30%. A result that needs another round is a call that did not bring enough, and that is where the work still is.
One change showed the cost of chasing. When an unanswered call waited an hour instead of ten minutes before following me into a chat, the daily chasing fell by 86%. Most decisions can wait an hour. So the numbers say: call a person less often, and bring more with each call.
I learned what a call must bring from one that failed. To answer one question, I had to search a whole planning document for the section it referred to. I did not find it. The call cost me more focus than the decision did.
I have cut social media almost to zero, and I noticed the agent chat pull at me the same way. It brought anxiety, not calm, and a fear of taking my attention away. That is my own observation, not research. Now I react only to changes. A message that needs nothing from me should not reach me at all.
Five rules hold the agents to this.
- A finished result comes back to me in the same task, marked for review.
- Work blocked on me becomes a task in my task list, not a reminder in a chat.
- My task list holds only what waits for me. The agents’ own work lives elsewhere.
- My phone hears from the agents in two cases only: a result to review, or work waiting for my action.
- Whoever finishes a piece of work cleans up after it.
Together they keep a call from reaching me unless it is a decision, and put everything I need in one place when it is, so there is less to rework, nothing to search for, and no feed to watch.
Try it on one stream of agent work for a week. Count the calls that reached a person, and the rounds each took before it was settled. A call that was not a decision should have stayed quiet. An extra round marks something the call should have brought.
Written by Pavel Khodalev, author of RACE Programming and CTO of First Line Software. Follow new essays via RSS.
