Anthropic has published a study of its own workforce, and the numbers are more specific than the usual commentary about AI and jobs because they are about the company doing the measuring.
By August 2026, interactive AI agent hours ran at roughly six times employee hours, counting human hours as 40 per week per head. Across all agent work, autonomous and interactive combined, agent hours first passed human hours in March 2026.
Two different junior numbers, often confused
The report gives two figures and they measure different things:
- The junior share of technical department employment fell from 21% to 14% since early 2025.
- The junior share of technical hires fell from 27% to 9%.
The second is the sharper number and the one worth watching, because hiring leads employment. Junior technical postings dropped sharply after December 2025.
Output, and whether it holds up
Code commits per engineer are about 5x their level six months earlier. Anthropic compared 192 engineers who adopted coding agents against 74 who did not, and found output rose sharply after adoption for the adopters.
The finding that matters most is the one about quality: agent-written and human-written lines were later removed or reverted at the same rate. That is the check most "productivity up 5x" claims never run, and it is the difference between more code and more working code.
What to take from it, and what not to
This is one company, and an unusual one — an AI lab with every incentive to use its own tools heavily and some incentive to publish a flattering result. The revert-rate finding cuts against the flattering reading, which makes the whole thing more credible than it would otherwise be.
What does generalise is the shape: the entry-level rung is where this lands first, because that is the work agents do adequately. The collapse from 27% to 9% in hiring happened in roughly eighteen months, at a company that was expanding.
Sources: Anthropic