Anthropic just published the future of work — and it's its own org chart

The org chart of the future, published by the company living it

Anthropic's internal research shows AI agents at the company working more hours than human employees since March 2026, and by August 2026 interactive agent hours outnumbered human hours six to one. Productivity rose with it: code commits per engineer increased fivefold in the six months preceding October 2026, and the company reports gains in code production, sales, and contract processing. The bottleneck has moved. Human attention, the company says, is now the primary constraint on output — AI acceleration made oversight, not execution, the scarce resource. m:…0794 m:…5687

The hiring mix changed underneath the productivity. The proportion of junior technical hires at Anthropic fell from 27% to 9% since early 2025, and the share of junior-level employees in technical departments dropped from 21% to 14%. Hiring managers now emphasize applicant autonomy and judgment over specific technical domain expertise. Meta-work — agent management and pipeline monitoring — grew from 10% of interactive Claude sessions in fall 2025 to 22% in spring 2026. This is a workforce recomposing itself around supervision: fewer people learning the craft by doing it, more people hired to watch machines do it. m:…0794

The expectations chasm

Anthropic researchers surveyed US adults and their own staff, and the two groups live in different decades. Anthropic staff expect AI to perform 75% of their current work by 2030; US workers put the figure at 28%, and only 5% expect 80% or more. On the bolder question — a Nobel-level AI discovery by 2030 — 91% of Anthropic staff say yes versus 39% of US adults. Both groups agree that people who expect AI to do more of their own work forecast higher unemployment, and both share similar economic models of AI's impact. The disagreement is only about speed. The people closest to the frontier see it arriving years earlier than the public does — which, for a workforce planner, is itself the signal: the adoption curve is set by the fast group, not the median one. m:…8741

The lab that proves its own forecast

The same week, Anthropic showed why its staff answers that way. Using Claude, researchers identified the origins of the NIST ECDSA curve seeds — a decades-old cryptographic mystery — in under half an hour. The seeds were generated with the C library's lrand48 generator seeded with a Unix timestamp, a simple reproducible method that quiets backdoor concerns; the discovery is explicitly not a cryptographic break. The $36,864 bounty is being donated to the Internet Security Research Group. This is not a demo: it is the kind of painstaking forensic work that used to take specialists months, done in an afternoon. m:…1278

In a separate result, Claude Mythos 5 worked in a closed loop with an automated evaluator to optimize eight quantum circuit primitives for chemistry, optimization, physics, and number theory — without human input, improving on the best published circuits by an order of magnitude or more in some cases. The FeMoco ground-state energy computation gate estimate fell from 50 trillion to 340 million. Its elliptic-curve point addition design uses 12% more qubits but 17 times fewer gates than the smallest previous design, completed in a single 53-hour run. Read against the labor data, the research week reads as the forecast and the evidence in one package: the machine does the specialist work, the humans manage the loop, and both sides of that equation are accelerating. m:…1649

The decision for executives

Anthropic generates two to four times as much revenue per employee as Meta and Google did during their own rapid growth phases, the company says, and it has built an early warning system to track jobs at risk of AI-driven disruption — plus internal redeployment strategies, with tenured employees moving departments 50% more often than last year. The lesson for executives is not to copy the numbers but to copy the instrumentation: measure agent hours against human hours, watch the junior pipeline, and name the bottleneck explicitly. Anthropic's data says the bottleneck is human attention. Most companies have not yet asked the question. m:…0794