Anthropic: agent hours hit 6-to-1 over humans, junior hiring collapses

The latest:

Agent working hours overtook human working hours inside Anthropic in March 2026 and reached a ratio of 6 to 1 company-wide by August, according to a report published by the Anthropic institute on 16 October 2026. In research departments the ratio hit 19 to 1. Over the same stretch, the share of junior hires in technical departments fell from 27% to 9%.

Details:

• The report: The paper, titled How employment is changing at Anthropic, was written by Nathan Wilmers, Regina Katipone, Maxim Massenkoff and Peter McCrory. It draws on internal employee data, Claude usage, code output, surveys, Slack and calendar records, hiring scorecards, and 17 staff interviews. Claude-based raters coded agent logs, task difficulty and scorecard comments.
• The ratio: Agent hours passed human hours in March 2026, the report says, and stood at 6 to 1 company-wide by August in employee-directed sessions, with research departments at 19 to 1. Roughly 80% of interactive Claude session time is spent waiting on human input.
• Entry-level squeeze: Junior staff fell from 21% to 14% of technical employees since the start of 2025, while the junior share of technical hiring dropped from 27% to 9%. Open entry-level roles declined sharply from December 2025, falling to a 6% share. Non-technical departments stayed broadly stable.
• Hiring mix: Technical departments’ share of total hiring slid from 61% to 56% over the past year, according to the report — a shift that runs alongside, not against, the broader agent rollout described in the same data.
• Productivity: Code commits per engineer are running about 5 times their level six months earlier. A staggered-adoption design compared 192 engineers who adopted coding agents with 74 who did not. The Decoder, citing the report, listed gains of 5.1x for engineers’ commits and 3.1x for data scientists’ queries.
• Beyond engineering: The same figures put legal at 3.1x in contracts processed, sales at 1.7x in closed deals, recruiting at 1.4x in hires and accounting at 1.4x in invoices. Non-technical staff went from 2 coding-agent sessions a month in August 2025 to 90 a month by May 2026.
• Meetings up: The median employee now spends 20% more time in meetings than a year earlier, with a 16% rise for individual contributors in technical roles. The authors link this to growing coordination work as execution tasks shift to agents.
• Intermediate work: The share of what the report calls intermediate work — managing agents and building tools — rose from 10% to 22% of sessions, marking a change in what employees do rather than how much they produce. Agent-written code showed no difference from code the same person wrote directly in the same month in how often it was later deleted or reverted.
• The caveats: Adoption was not random, so selection bias is possible, the report says. It presents measured data rather than forecasts, and notes Anthropic is a special case as an AI-intensive software company, meaning results may not carry to the wider economy. Rapid growth and voluntary adoption may also distort the numbers.
• Companion paper: A second report published the same day, The AI expectations gap, found more than half of Anthropic employees expect AI to do at least 80% of their current job by 2030, against 5% of US workers. It said 91% of employees expect a Nobel-level discovery from AI by 2030.

Between the lines:

The internal numbers point in two directions at once. Output per worker rose across every function measured, yet the biggest structural shift falls on entry-level technical roles, whose hiring share dropped to 9%. The parallel rise in meetings and in agent-management work suggests the surviving job is becoming coordination rather than execution — an inference the report’s own authors draw from their meeting data.

What’s next

Watch whether Anthropic’s junior hiring share recovers from 6% of open roles, whether other frontier labs publish comparable internal data, and whether the staggered-adoption design is replicated at firms outside AI-intensive software.