once agents do the work, a company's bottleneck stops being labor and becomes the attention of its humans. how much work a unit of attention can direct is now the number that separates companies, and the gap is enormous: the best operators direct tens of times more than the median. most teams optimize for how much the ai can do. the real lever is how little it can bother a human without going silent.
agAman Gour // FurtherAI@amangour30A friend recently joined Anthropic as an engineer.
He told me:
1. Everything he does is being “watched.”
2. AI proactively surfaces internal discussions where he can contribute.
3. It even suggests what he should say.
That’s fascinating.
How Anthropic uses AI internally might be an early glimpse of how every enterprise will operate.— Aman Gour // FurtherAI (@amangour30) August 19, 2026
take the post at its word and read it as an org chart: software deciding where a person's attention goes and pre-drafting what they contribute. the same company reports that by may 2026 more than 80% of the code merged into its codebase was written by its model, up from low single digits in february 2025, and that between february and august 2025 its coding agent went from 9.8 to 21.2 actions between human check-ins while human turns per task fell from 6.2 to 4.1.
put those together and the company is no longer short of hands. what it has a fixed amount of is the attention of the people who decide what the hands do. when agents do the work, the bottleneck moves from labor to attention, and the operating model becomes a routing problem for one person's judgment.
herbert simon saw the shape in 1971: "a wealth of information creates a poverty of attention." for fifty years that was a complaint about inboxes. now it is the production function. a company's output used to be people times hours. with agents, output is attention times multiplier: how much agent work one unit of human judgment can direct before it has to be spent again.
we are not the only ones reading it this way. the same week, a widely shared thread on x put it in exactly these terms, citing openai co-founder greg brockman on its coding agent going from writing 20% of the company's code to 80% in four months:
bbbroke boy@0xfuckpovertyOpenAI co-founder Greg Brockman says AI coding tools went from writing 20% of code to 80%.
In four months.
The bottleneck is no longer code.
It is human attention.— broke boy (@0xfuckpoverty) August 2026
the hiring consequence follows directly, and several of these threads land on it: teams will stop hiring people primarily to produce code and start hiring people who can direct large numbers of agents. what they will be hiring, in other words, is attention worth multiplying.
the attention side is fixed. a person has a few hours a day of real judgment and it does not scale. so everything that changes a company's output now lives in the multiplier, and the multiplier is an engineering variable: it is set by how well the system decides what reaches a human, in what shape, and what never should.
two published numbers bound it today. at the frontier, 21.2 agent actions per 4.1 human turns is about 5 agent actions per human touch on a task, measured at anthropic in august 2025. in an agent-native company built for routing, we measure 47.6 agent actions per human message over thirty days. same idea, different denominators, an order of magnitude apart.
human output was never a bell curve. the largest study of individual performance (633,263 people across 198 samples) found it follows a power law: a long tail of people producing most of the result. the multiplier sits on top of that and stretches it. the honest pre-ai spread between the best and median engineer is about 10x. with agents, the spread between someone using a chat window and someone directing a routed fleet is measured in orders of magnitude, and the gap widens faster at the top than anywhere else.
where companies sit on it is mostly a matter of claims, and the claims run hot: dora reports that across 425 organizations with instrumented repositories, measured ai-authored code was 22% in late 2025 and 27% in early 2026, while 34% of organizations said theirs was over 60%. read every executive "share of our code" number against that 2 to 3x haircut. two companies have published numbers that survive it, and both build frontier models. anthropic's 5.2:1 mark on the chart is august 2025; by may 2026 it reported more than 80% of merged code written by its model, up from low single digits fifteen months earlier, so its number today is well above anything drawn here. openai reports 97.9% of employees on its coding agent and 99.8% of internal output tokens through it. we read that as causal rather than coincidental: more output per person is the advantage in a race decided per quarter. mckinsey is reported to find roughly 6% of companies are high performers on ai, with at most one in ten scaling agents in any single function.
metr measures how long a task an agent can finish: at 50% reliability, the frontier is hours and doubling every few months. at 80% reliability it is about 27 minutes, and that number has barely moved release over release. reliability, not capability, decides how often a human has to be touched.
everything an agent does reaches a person in one of three shapes: it decided and said nothing, it decided and left a note, or it stopped and asked. the multiplier is set by how much of the work lands in the first two shapes without anything important going silent.
companies miss in both directions. over-asking is invisible to the people paying for it: in metr's 2025 randomized trial, experienced developers given ai tools were 19% slower on their own repositories and believed they had been 20% faster. under-asking is visible only after it fails: klarna said in 2024 its assistant was doing the work of roughly 700 agents, and by 2025 its ceo was saying cost had been too dominant a factor and quality had dropped; it rehired people and built an explicit hand-off to a human. the routing was what failed.
we measured our own routing to see where the waste lives. it turned out to be in the questions: most of what stopped a human for an answer should have been a note.
seven in ten questions should have been notes. the fix was rules, an instrument that reads every outgoing message, and a ledger for promises an agent made and dropped; the audit now runs weekly and tightens the filter each pass. the multiplier went up without touching a model.
if you work with agents, your attention is the denominator of everything you produce. the goal to optimize is how little the ai has to bother you without going silent, and getting there is a re-orchestration job, done with rules and routing rather than a better model.
re-orchestrate the workflow, before upgrading the model. an agent-native process is designed backwards from your attention: what may run silent, what arrives as a note, what genuinely stops you. reliability horizons move slowly; a routing filter moves today.
measure your number, then raise it by routing. agent actions per human touch, and the share of touches that were real decisions. give every message a state tag (e.g. "done — nothing needs you" / "driving" / "needs you: one decision"), filter every ask, and keep a short written list of what may never go silent. most people have never seen their own number; ours rose without retraining anything.
this post took about 18 minutes of one human's time, dictated. everything else was agents.