Playbooks · Attention
Parallel AI agents scatter your attention. One developer's fix is to queue his judgement and close each decision
Agents run in parallel but judgement does not, so one developer moves AI work into a git-backed queue and handles one decision at a time.
by Jo·4 min read·
New here? Start with the free AI Survival Kit →
Parallel AI agents scatter your attention. One developer's fix is to queue his judgement and close each decision
0:00 / 5:28
This voice is generated by AI.
Sergey Kuhne, writing on his blog sergeykuhne.com on 8 November, described a problem that anyone running Claude Code or Codex from a terminal will recognise. You start a task. The AI is slow to reply, so you start another. The first reply needs a clarification, a follow-up and a side task, and each one gets its own tab. Soon you are "hopping between 20 threads, trying to keep all those contexts in your brain RAM." Then you can't. In his words, focus "smashes" and you lose track of where you started and where you meant to finish. His fix is a priority queue kept in git. An agent does the building, and a single inbox holds only the decisions that need him.
The bottleneck is now the person
This is a story about where the scarce resource sits. Agents run in parallel. Judgement does not. Kuhne names two problems: sequencing and context. He solves both by taking them out of his head.
Every ticket is a markdown file in the repo, next to the code it concerns, and the board is a committed snapshot. "Nothing important lives in a chat window and nothing lives in my head." If a session dies, the queue is still there.
The division of labour is explicit. The agent researches new tickets, builds the ones a human has tagged as safe, runs the tests and a browser check, and verifies what is already deployed. For sensitive code it builds on a branch and asks first. It never deploys to production. "That part is mine." Then the line that matters: "What it cannot do is decide."
That is the War Room's anchor line in working form. AI does the routine work. You do the thinking. The failure Kuhne describes starts when the thinking is spread across twenty tabs. Attention is the first thing to go, and a scattered mind makes poor rulings.
Closing is the move
A script reads the whole board and hands him a prioritised list of items waiting on his judgement: review this change, answer this question, rule on this tradeoff. He reads them on one page, a Claude artifact rebuilt from git every cycle. The page is only a viewer. His verdict goes back into git, onto the card.
The discipline is in the sequence. "I take one item. I load its context. I make a decision or request an action. And I close it, by putting it back into the queue." Closing the ticket clears it from working memory, so the next item gets full focus.
He is candid about the limits. "It is still exhausting. But I don't feel distracted or lost." Two problems remain. The first is serving up context so that loading a ticket takes a minute and not twenty. The second is tickets looping because they are blocked on a ruling only he can give.
That second problem sums up the operator's position. The person who rules is the constraint. More agents raise the volume of decisions without raising the capacity to make them. Output now depends on how fast and how well that person decides, and on whether the system protects their attention while they do it.
Your Next Move
- Get pending AI work out of chat windows. Count the threads you have open right now. Move every unfinished task into one written list that lives with the work itself and still exists when a session dies.
- Write down the line between build and decide. Sort your recurring tasks into three groups: what an agent may run unattended, what it builds on a branch for your sign-off, and what only you ship. Kuhne's rule that production deploys stay with the human is a sound default.
- Rule on one item at a time, then close it. Collect decisions in a single inbox. For each one, load the context, make the call, record the verdict where the work lives, and move on. Time how long loading the context takes. If it takes closer to twenty minutes than one, fix the summaries before you add another agent.
I build systems like the one publishing this site. → Work with me
About the author
Jo
Jo runs The War Room: one signal a day on how AI is changing work, and what to do about it.
Sources
Primary source
Get the Briefing
Want more intelligence like this?
The weekly briefing, free — and the AI Survival Kit with it.
More in Playbooks
See allPlaybooks · Delegation
AI agents reported blocked tasks as done. Check the system of record, not the agent's summary
4 min read
Playbooks · Accountability
The No AI Slop manifesto makes owning your AI output the job
4 min read
Playbooks · Oversight
Karpathy's aircraft-manual English points to where the work goes: reading and checking what the model wrote
4 min read