Agents work six hours for every human hour at Anthropic, and attention is the limit

AnthropicPVT$350B — Anthropic, private, latest valuation $350B has turned its own payroll into evidence. On October 16 the company behind the Claude models published a study built on its employment records, agent telemetry, output metrics and 17 staff interviews↗ ¹. By August, agents run by its own employees were logging six hours of work for every hour those employees worked ¹.

The study lands ahead of a share sale expected to be the largest IPO on record↗ ² and now slated for November ³⁵, with some investors discussing a valuation near $1.5 trillion↗ rather than the roughly $2 trillion once expected ³⁵. It speaks to the question that price rests on: whether a lab's output can keep growing faster than its headcount. On this evidence it can, and the binding constraint has moved from the model to the person watching it.

Engineers who adopted agents added far more code than those who had not

The strongest evidence is an internal comparison: engineers who adopted coding agents set against colleagues in the same month who had not. Adopters added far more code in the months that followed, and their agent-written code was no more likely to be deleted or reverted later ¹. The report supplies its own caveat: adoption was not random, and the engineers who adopted were already ramping up output faster than everyone else ¹.

Code velocity per engineer has risen roughly fivefold since spring ¹. At five years old, Anthropic books two to four times the revenue per employee that Meta▼−4.28% — Meta, down 4.28 percent today and Google managed at the same stage ¹. Headcount still tripled in the past year ¹; the mix changed, and the entry rung is what got cut.

Human attention is the constraint the report names

The same telemetry caps the story. Interactive agent sessions sit idle about four-fifths of the time, waiting for a person ¹. Reading everything the agents produce would take more than half the work week, even for non-technical staff ¹. One engineer described facing "an army of Claudes, with every one asking for my attention" ¹.

So the work rebuilt itself around supervision. Non-technical staff went from two coding-agent sessions a month in August 2025 to 90 by May 2026 ¹. Meta-work such as writing agent specifications and monitoring pipelines roughly doubled its share of sessions, and hiring managers now screen for autonomy, judgment and a track record with AI tools over narrow domain expertise ¹.

The cost is quieter than a layoff

Outside the lab, the jobs debate still runs on layoff tallies↗ ⁴. This study is telemetry from inside the machine, and the cost it shows is quieter than a layoff: the rung where people learn the job is disappearing, and the report is candid that redeploying incumbents does nothing for those never hired ¹.

For the listing itself, the wage bill is not where the risk sits. The $518 billion of committed compute behind the models is ³⁵.

The skill in demand is directing agents

The signal for everyone else is already in the hiring bar. The skill in demand is directing agents: scoping work you can verify, writing standing instructions, judging output. It is learnable, and this report is the best argument yet for starting today.

Whoever builds the layer that cuts the human attention each agent-hour needs is likely the one turning capability into output, and that is likely where the value migrates.