Anthropic's Own Data: Agent Hours Outnumber Staff 6 to 1

§ 01Agents now log more hours than people#

Based on agent session telemetry, the sum of autonomous and interactive agent work hours passed human work hours for the first time in March 2026. Counting only the interactive agents individual employees direct, with subagents pooled, agents worked six times as many hours as their human handlers by August. Agent time is runtime including inference and tool use, excluding any wait longer than 30 minutes; human hours are headcount times 40 a week.

Output moved with it. Code commits per engineer are 5x their pace of six months earlier, and lawyers processed more contracts, accounts payable cleared more invoices and salespeople closed more deals. To separate AI from the company’s own boom, the researchers followed 192 engineers who adopted coding agents against 74 who had not yet, and the adopters’ added code rose sharply after adoption. Agent-written code was removed or reverted no more often than code the same person wrote by hand, in the same month and directory. Anthropic adds two cautions: adoption was not random, and more code does not necessarily mean better code.

§ 02Fewer junior technical roles#

Talk of lower demand for junior technical staff started inside Anthropic’s core technical teams in January 2026. By April, engineering and research had de-emphasized hiring for lower-band technical roles, and about a third of ML managers and a fifth of software engineering managers said they were hiring less because of AI.

This is a mix change, not a hiring freeze. Anthropic tripled in size in the last year, technical hiring still grows in absolute terms, and sales, finance and recruiting hiring has grown faster. At five years old it takes in 2 to 4 times as much revenue per employee as Meta and Google did at the same age. Clerical roles stayed thin as the company grew, and employees with a year’s tenure were 50% more likely to move across departments each month in 2026 than in 2025. A product writing team that once wrote the greetings on Claude.ai now writes model evals and prompts: “Writers who used language to drive human behavior are now using it to drive Claude behavior.”

§ 03Human attention is the bottleneck#

In interactive sessions, about 80% of an agent’s time is spent waiting for human input; as one engineer put it, people are “bottlenecking the agents.” Converting output tokens to words at 250 words a minute, reading everything their interactive sessions produce would take non-technical groups more than half the work week. Another engineer described the job as facing “an army of Claudes, with every one asking for my attention.”

The response is meta-work: building the tools to steer and check agents, which rose from 10% to 22% of interactive Claude sessions. Written standing instructions for agents grew from a few hundred to 8,500 in a year. One engineer set up an agent chief of staff to triage agent requests; another wrote a performance review for a team of Claudes. For non-technical staff, the strategic part of the job grows: “that 20% that was strategic becomes 90% of the job.”

People did not disappear from the loop. The median employee spends 20% more time in meetings than a year ago, and 16% more among technical individual contributors. Agents appeared in nearly half of active public Slack channels but added replies rather than replacing human ones. They also forced a habit: decisions made in unlogged meetings leave agents working from stale context, so work at Anthropic moved into the open. One employee whose name sat on an old swag document kept getting agent requests for t-shirts long after the job had moved on.

§ 04What hiring managers look for now#

Across thousands of interview scorecards since 2024, mentions of a candidate’s track record with AI tools tripled for technical roles (from around 5%) and quadrupled for non-technical roles (from around 7%). For technical candidates, the biggest gains were comments on autonomy and judgment, while domain expertise shrank as a share of assessment. For non-technical candidates, customer and mission orientation rose and execution skills fell.

§ 05The expectations gap#

The second report asks whether Anthropic staff and the public disagree about how AI affects jobs, or about what AI can do. Its answer is the second.

Among both US workers and Anthropic staff, people who expect AI to do more of their own job anticipate higher unemployment, and the link is very similar in both groups. The authors read the gap as different experiences of what AI can do, not different theories of how it changes work.

§ 06What Anthropic says employers can do#

The report names three practices. First, widespread stock ownership, so employees share in the gains from rollout; it calls a way to share the benefits essential while noting that shared ownership may not suit every firm. Second, an early warning system that reads role-level demand and productivity against AI capability and use. Third, internal redeployment of employees whose work is changing. Anthropic is direct that none of this prevents disruption or helps people who were never hired, and calls for government policy on retraining and income support alongside it. In one executive’s words, “a workforce that is anxiously worried is not a workforce doing its best work.”

§ 07What it means for teams running agents#

The most practical findings are organizational. Agents work best from fresh, written context. Individual contributors end up managing a team of agents, and they spend real time building the tools to supervise them. That is the design we built CellCog around: AI employees post in shared channels, keep their work on a task board, write a handover so the next session starts from the record, and bring a person only the decisions that need one. When one needs help, it can hand work to helper agents and review every result before calling it done.