Anthropic reports agent hours at six times employee hours, with humans as the bottleneck

The report draws on Anthropic's internal employment, usage, output, survey, calendar and hiring records, plus 17 employee interviews. Its headline figure is that by August, interactive agent hours outnumbered employees' hours six to one. Sessions are mostly idle: the authors estimate Claude sessions spend around 80% of their time waiting for human input, and that reading all the output of their interactive sessions would take even non-technical employees over half a work week. One engineer is quoted describing "an army of Claudes, with every one asking for my attention."

The response the report documents is a change in the work itself. Meta-work, meaning skills, plans, reviews and monitors for other agents, rose from 10% of the average employee's interactive sessions in fall 2025 to 22% in spring 2026, and written standing instructions for agents grew from a few hundred in July 2025 to 8,500 a year later.

On quality, the authors compared code written by the same person, in the same month and directory, by hand versus by a coding agent, and report no difference in the rate of lines later removed or reverted. They also followed 192 engineers who adopted coding agents against 74 who did not, and report added code rising sharply after adoption. Both results carry the authors' own caveats: adoption was not random, engineers who moved to fully autonomous agents were already on a steep upward trajectory, and more code is not necessarily better code. Junior technical hiring fell, with the junior share of technical hires dropping from 27% to 9%. The figures are Anthropic's, about Anthropic, and we could not see the underlying data. We read the report text but not its charts.

Why it matters: For anyone scaling agents, this is first-party evidence from a heavy user that review attention, not agent throughput, becomes the constraint, and that the response is building supervision tooling rather than reading more output. The revert-rate result is a single company's proxy for quality, so treat it as a data point for your own verification loop, not a general finding.