The Constraint Moved

At Anthropic, agents now log six times the hours of the people who direct them, and the company's own study names the new bottleneck: human attention. Goldratt would not be surprised. What changes when the scarce resource is judgement, and why most organisations are not there yet.

Thomas Forwe

Thomas Forwe

— 6 min read
Share

On 16 October 2026 Anthropic published a study of its own workforce, How employment is changing at Anthropic. One sentence in its key findings is worth reading slowly: “By August interactive AI agent hours outnumbered employees' six to one.”

The easy reading is that the company has become six times more productive. The study does not say that, and it would be the wrong lesson. The more interesting finding sits two sentences further on: “This acceleration makes human attention the bottleneck for agent output.”

The constraint did not disappear. It moved.

What the study measured

The study is unusually explicit about its data, which makes it worth taking seriously. It draws on internal records covering employment, AI use, output metrics, employee surveys, communication patterns, calendar data and hiring decisions, and it adds 17 employee interviews. Four of its observations carry the argument here.

Agent hours. The six-to-one figure counts only interactive sessions directed by employees, with subagents pooled. Agent time is runtime, “including inference and active tool use, but excluding any wait longer than 30 minutes”. Human time is headcount multiplied by an assumed 40-hour week. This is a measure of volume, not of value, and the authors do not present it as anything else. Counting autonomous agents as well, agent hours passed human hours for the first time in March 2026.

Idle time. In interactive sessions, most of an agent's time is not spent working. “Claude sessions amount to around 80% idly waiting for human input.” One engineer's description: people are “bottlenecking the agents.”

Reading time. The authors estimate how long it would take to read everything agents produce in interactive sessions, converting tokens to words and assuming 250 words read per minute. “Even among non-technical employees, reading all the output of their Claude agents would take half their work week.” That is an estimate, but the direction is hard to argue with.

Meta-work and hiring. Employees respond by building visibility tools and monitoring agent pipelines. The share of interactive sessions spent on this kind of meta-work rose from 10 percent in autumn 2025 to 22 percent in spring 2026. Written standing instructions for agents grew from a few hundred in July 2025 to 8,500 a year later. Hiring managers “increasingly emphasize autonomy and judgment, relative to technical domain expertise and implementation.” And the share of junior employees in technical departments fell from 21 to 14 percent since early 2025, junior technical hires from 27 to 9 percent.

One engineer described the working day as facing “an army of Claudes, with every one asking for my attention.”

Goldratt would recognise this

None of this would have surprised Eliyahu Goldratt. The ideas he set out in The Goal in 1984, later called the Theory of Constraints, rest on an observation that sounds trivial and is routinely ignored: every system has one constraint that limits its output, and speeding up anything else does not raise output. It only builds inventory in front of the constraint. When you do relieve the constraint, it does not vanish. It moves somewhere else.

That is what the Anthropic data shows. For most of the history of knowledge work, producing the work was the constraint: writing the code, drafting the contract, preparing the analysis. Agents have relieved that step so thoroughly that the constraint has moved downstream, to the person who has to judge whether the output is right, decide what happens next and take responsibility for it. The inventory is visible too: sessions waiting for input, output nobody has time to read.

Goldratt later condensed what to do next into five focusing steps. They translate almost directly.

  1. Identify the constraint. It is no longer production capacity. It is review, decision and accountability, which in the end means judgement. Most AI business cases still count hours of production saved. They are measuring the step that has stopped being scarce.
  2. Exploit it. Do not waste the constraint's time. Every minute a reviewer spends reading output that a check could have rejected, or reconstructing context that the agent could have summarised, is a minute lost for the whole system. Goldratt's line was that an hour lost at the bottleneck is an hour lost for the system. Verification before human review, short and comparable summaries, and a clear statement of what the reviewer is being asked to decide are not conveniences. They are throughput.
  3. Subordinate everything else to it. Starting more agents than people can judge does not raise output; it raises the queue. The number of parallel agents should follow review capacity, not token budgets. In lean terms: a limit on work in progress, set at the constraint.
  4. Elevate it. Add judgement capacity. This is where the Anthropic findings become uncomfortable, and I come back to it below.
  5. Do not let inertia become the constraint. Once review is organised, the constraint will move again, perhaps to the quality of specifications, to the data agents depend on, or to the decision rights nobody wrote down. The organisation has to keep looking.

The uncomfortable part: where judgement comes from

Judgement is not a skill you hire once. It is built by doing the work. You learn to spot a bad contract clause by having drafted contracts, and a fragile piece of code by having written and broken some.

The study observes both sides of this. On one side, employees “are increasingly removed from the object-level work that originally built their skills.” On the other, the share of junior technical hires fell from 27 to 9 percent. Read through Goldratt, the resource that has become the constraint is the one whose supply is being reduced at the entry point. An organisation can live with that for a while on the judgement its experienced people already have. It cannot live with it indefinitely.

I do not read this as a criticism of Anthropic. The study reports it openly, and the company describes an early warning system for jobs at risk and a preference for internal redeployment. But anyone who plans to elevate the constraint has to answer a question that most AI strategies do not ask yet: where will the next generation of reviewers learn to review?

Most organisations are not there yet

Two days before the Anthropic study, on 14 October 2026, the European Commission's Joint Research Centre published figures on AI adoption in European firms, based on Eurostat's survey of ICT use in enterprises. “The share of firms using AI in EU strategic sectors grew from 7% in 2021 to 18% in 2025”, and across the wider economy from 8 to 20 percent.

The growth was “driven mainly by accessible, general-purpose AI — systems that analyse or generate text, code, images, sound and video.” What lags is depth: “in many cases AI still needs to be embedded into core business processes, such as production.” And one finding points in the same direction as everything above: “The firms investing more in structural organisational changes to integrate AI are more likely to use AI in production.”

That is a correlation across sectors, not proof of cause, and the data end in 2025. But it describes a different position on the same curve. For most firms the constraint is not yet human attention. It is the step before: AI is used, but not built into how the work is organised. Tools are not the scarce part. The capacity to redesign processes, roles and decision rights around them is.

Put the two studies side by side and the sequence becomes visible. First the constraint is integration. Once integration succeeds, it moves to judgement. Organisations that plan only for the first step will meet the second unprepared.

Three questions for your own organisation

Usage adds output. Transformation decides where judgement sits. Three questions make the difference concrete:

  • Where does agent output already wait for a person, and how long? If nobody measures that queue, nobody knows where the constraint is.
  • Who decides, and on what basis? Review is only fast when the reviewer knows what they are entitled to decide and what they are not. Unclear decision rights turn every output into an escalation.
  • Where do the next reviewers come from? If agents take over the work that used to train people, something else has to train them.

What this does not show

Three limits, so that the argument does not grow beyond its evidence. The Anthropic study describes one company, written by that company, at the frontier of AI use; it is an observational study of internal data and interviews, not an experiment. Agent hours are not productivity. And the JRC figures count firms that use AI, not how well they use it.

What remains is a pattern that is older than any of these models: when one step of a system speeds up, the constraint moves. This time it has moved to the one resource that cannot be scaled by adding compute.

Sources

New posts by email: subscribe to the newsletter. One email per post, no tracking.

Read more