Anthropic reports: agent hours now exceed human hours; staff say AI will do 75% of their job in 4 years

新智元 · wechat · 2026-10-17

Anthropic released two internal reports showing a widening gap between insiders and everyone else.

Employment at Anthropic

In March, total agent hours at the company first exceeded human hours; five months later, employee-delegated agent hours hit 6x human, and 19x in research.
Per-engineer commits are up 5.1x; one Claude sweep cleared backlogged API issues with 800+ fixes (est. 4 years of human work); non-technical staff's monthly coding-agent sessions grew from 2 to 90.
Employees would need 424 hours/week to read all Claude output — Claude idles 80% of the time waiting on humans. Daily code changes exploded from 2,500 to 30,000+, breaking Git; 7 engineers spent 5 months building a replacement VCS (Scion), with a Claude-rewritten import tool using 1/20 the memory at 4x speed.

The expectations gap survey

1,334 employees estimate AI will do 75% of their current job within four years; American workers say 28%. 91% of staff believe Nobel-level AI discoveries arrive before 2030; only 39% of US adults agree.
Staff estimate 51% of frontier-model R&D is already done by AI (research teams: 60%), rising to 80% by end-2027 and 95% by 2030.
Junior share of new technical hires fell from 27% to 9%.
Anthropic's chief economist says the company leads the broader economy in AI adoption by about 18 months.

More from AGI Musings