Look at an org chart long enough and you see what it really is: a seating plan for attention.
Ten boxes, ten people, ten threads of attention. Every process underneath the chart is sized to those threads. An approval chain exists because attention is scarce and must be rationed. A weekly meeting exists because ten threads must synchronize. A headcount plan exists because threads cost salaries and arrive one per person.
None of this was ever written down as a law. It never had to be. For the entire history of organizations, one person meant one thread, and the thread tired, slept, and held one problem at a time. The desk assumption was so total it was invisible.
It is visible now, because it broke.
What broke, exactly
The machines did not make people faster. They made threads cheap and simultaneous, which is a different event entirely.
The capability numbers move monthly, so they carry dates. The length of task a system can complete unaided has been doubling roughly every seven months since 2019, measured by METR in March 2025, with newer frontier models possibly doubling faster. Meanwhile the human thread still reads at about 238 words a minute and still attends one meeting at a time.
So one person can now steer several threads of real work, not someday, now. Ten people steering ten agents each is a different unit of work, and the desk assumption has no box to draw it in.
What the assumption was quietly carrying
Because the assumption was invisible, a lot of machinery was resting on it without saying so:
- Headcount plans. An FTE is a bundle of threads. When threads decouple from people, planning in FTEs starts answering the wrong question.
- Spans of control. The old limit was how many people one manager can attend to. That limit was never about management. It was about attention arithmetic.
- Approval chains. Most approvals ration scarce attention. When attention is abundant at the working layer, an approval is no longer a checkpoint. It is a queue.
- Process design. Handoffs exist because a thread puts work down at five and picks it up at nine. A thread that runs overnight does not need the handoff. It needs a boundary and a log.
None of these are evil. They were correct arithmetic for a scarce resource. The resource changed price and supply, and the arithmetic did not.
What does not follow
The tempting conclusion is that ten threads produce ten times the work. They do not, and the reasons are the useful part.
Threads multiply output only where output was the constraint. Most mature organizations are not output-constrained. They are verification-constrained, decision-constrained, or trust-constrained. A team that cannot review last week's work has no use for ten times last week's work, and adding threads to that team just builds a bigger queue. The METR trial that found experienced developers taking 19 percent longer with AI tools on real repositories is the caution here, and it is dated, early 2025: speed added at the wrong node slows the system.
The other thing that does not follow is abandonment. The desk assumption was the load-bearing wall of a lot of quiet wisdom: handoffs force clarity, approvals force decisions, meetings force honesty. Break the wall without understanding what it carried and you get chaos with better tooling.
The questions worth asking
The assumption is worth revisiting question by question, not sweeping away:
- Which of our processes exist to ration attention, and what changes if attention stops being scarce at the working layer?
- Where would ten parallel threads of real work change what we attempt, not just how fast we attempt it?
- Where is our actual constraint this quarter: output, verification, decision, or trust?
- If a thread can run all night, what does the overnight shift own, and who reads its log at nine?
The last question is the uncomfortable one, because a thread that runs overnight also needs someone who owns what it did. That is a management question, and it does not automate.
FAQ
Is this just saying AI lets everyone do more?
No. Doing more keeps the unit of work fixed and adds speed. The desk assumption broke because the unit changed: from one attention thread per person to many steered threads per person. Organizations are built from units of work, so the chart, the plans, and the processes all inherit the change whether they admit it or not.
Does this mean headcount planning is dead?
Headcount still decides judgment capacity, ownership, and who answers when something goes wrong. What stops making sense is using headcount as a proxy for throughput. A plan that says ten people and a plan that says a hundred steered threads are different plans, and only one of them describes the working system.
Our processes work fine. Why change them?
They work fine because they are still rationing something. The check is simple: find one approval, one handoff, or one recurring meeting and ask what would break if a competent agent did the work overnight and a named human read the log in the morning. If the answer is "nothing, but we never tried," the process is rationing habit, not attention.
Where should a small team start?
With one overnight thread. Pick one bounded, reversible piece of work, give it to an agent on a written contract, and let it run after hours for two weeks. Read the log every morning. The desk assumption will feel less like a theory after the first week of reading what happened while nobody was at the desk.
Related reading
- Ten People, One Hundred Agents
- From Pull Requests to Proof
- The Review Bottleneck Is the Real Cost Center in AI Teams
- What Breaks Under Scale Was Usually Present at Small Size
- AI Needs a Harness Before It Needs Autonomy
- The New Venture Stack: Small Teams and Agent Fleets
If you are ready to plan in threads instead of desks, book a discovery call.