The headline number
Anthropic published a study of its own workforce on 16 October, and the figure it leads with is a ratio. Agent runtime first exceeded the hours its employees work in March 2026. By August, the report says, agents were logging roughly six hours for every hour worked by the people directing them.
The company counts agent runtime as inference plus tool use, excluding any wait longer than 30 minutes, with subagents pooled into the session that spawned them. Human hours are assumed at 40 a week per head. Only sessions an employee directed are counted, so the ratio measures delegated machine time against the time of the people delegating it, not the whole company.
Junior hiring
The second set of numbers concerns who gets hired. Since early 2025 the junior share of Anthropic’s technical employees has fallen from 21% to 14%. The junior share of technical hires has fallen further and faster, from 27% to 9%, with junior postings dropping sharply after December 2025. Over the past year the technical departments’ share of total employment has slipped from 61% to 56%, in a company that has roughly tripled in size over the same period. Anthropic does not give an absolute headcount.
The hiring finding comes from automatically coding interview scorecards for thousands of hires since 2024. Over that period, mentions of a candidate’s use of AI tools rose about threefold in technical roles and fourfold in non-technical ones.
What people inside the company are doing instead
Agent use has spread past the engineering organisation. Non-technical staff ran about two agent sessions a month in August 2025 and about 90 in May 2026. Within interactive sessions, roughly 80% of elapsed time is the model waiting for a person to respond — the constraint, on Anthropic’s own account, is human attention rather than compute.
The shape of the remaining work is changing too. Meta-work, meaning time spent setting up, steering and reviewing rather than producing, rose from 10% of interactive sessions in autumn 2025 to 22% in spring 2026.
The caveats the report carries
Anthropic is direct about what its numbers cannot support. Tool adoption was not random: the productivity estimate compares 192 engineers who adopted coding agents with 74 who did not, using person and month fixed effects, and the report notes that engineers planning large projects may have adopted the tools for that reason. Those who took up fully autonomous agents were already on a steeper output trajectory before they did.
Code quality is measured by deletion and revert rates, comparing a person’s own code with the code their agent produced in the same month and directory — proxies that, the report says, may miss errors specific to AI-written code. More code does not mean better code or more business impact. One survey item was reworded in April 2026, so that year’s answer is not directly comparable. And the whole study describes a single AI-forward company that sells the tools in question, which Anthropic says may not generalise to the wider economy.
What to watch
The junior hiring series is the one with consequences outside the building. If other firms publish comparable figures, the question becomes whether a 27-to-9 collapse in junior technical hiring is a property of this company or of the job. Anthropic’s internal data runs through August 2026, with some series ending in April, May or July, so the next update is the test of whether the ratio keeps climbing.
