Anthropic: agents now log 6x the hours of its staff

What the internal data shows

How employment is changing at Anthropic draws on internal data on employment, AI use, output and hiring, plus 17 interviews (Anthropic). Key numbers:

• In technical departments, the junior share of headcount fell from 21% to 14% since early 2025, and the junior share of hires from 27% to 9%.
• Code commits per engineer are five times the level of six months ago. Comparing 192 engineers who adopted coding agents with 74 who did not, code added rose sharply after adoption, and agent-written code was not removed or reverted more often.
• Agent work hours first exceeded human work hours in March 2026. By August, interactive agents launched by employees worked six times as many hours as their humans; in research organisations the multiple was 19.
• Anthropic brings in 2–4 times as much revenue per employee as Meta and Google did at the same age.

Source: Anthropic. Agent hours per human work-hour, February to August 2026, log scale.

Human attention has become the bottleneck. About 80% of interactive agent session time is spent waiting for human input. Even for non-technical staff, reading all of Claude's output would take half the work week. Employees respond with “meta-work”: writing specs for agents, having agents review other agents and monitoring pipelines. That share rose from 10% to 22% of sessions, and standing instructions for agents grew from a few hundred to 8,500 in a year.

Hiring criteria are shifting too. For technical candidates, autonomy and judgement now carry more weight relative to narrow domain expertise. Mentions of creative AI tool use in interview assessments have tripled for technical roles and quadrupled for non-technical roles since 2024.

Source: Anthropic. Junior share of hires, headcount and open roles in technical and non-technical departments.

What staff expect versus the public

The AI expectations gap survey was fielded by Morning Consult to 10,980 US adults from August 11 to 23; 1,334 Anthropic staff answered from August 9 to 12 (Anthropic).

• 91% of staff and 39% of US adults expect AI to make a Nobel-level discovery by 2030.
• On average, staff expect AI to be able to do 75% of their own work by 2030; US workers say 28%.
• The median staff estimate is that AI already does 51% of the work of building a frontier model, rising to 80% by end-2027 and 95% by 2030. These are personal views, not company forecasts.
• Run through an economic model, the median US adult's answers imply 5.4% GDP growth in 2030, the median staff member's 17.1%.

One notable match: in both groups, people who expect AI to take over more of their own job forecast higher unemployment. Those expecting 80% or more forecast 12.2% unemployment in 2030 on average. At the same time, 95% of staff expect Anthropic's headcount to keep growing.

How to read it

Both are Anthropic's own publications: the internal data describes a company that sells AI, and the staff answers are personal expectations. The company itself notes that agent adoption was not random, which can bias the productivity estimates.

The drop in junior technical hiring, however, is measured, and it may be the most important signal for other employers. Anthropic's response is an early-warning system that tracks roles at risk, plus internal redeployment: staff with at least a year's tenure were 50% more likely to move departments each month in 2026 than in 2025. The report concedes this does nothing for people who were never hired.

Key takeaways

• By August, agents at Anthropic worked six times the hours of the people directing them.
• The junior share of technical hires fell from 27% to 9%.
• Human attention is the bottleneck: agents spend about 80% of session time waiting for people.
• Staff expect much faster AI progress than the US public does.