Anthropic's agents outwork its people six to one. The scarce part is the person.
18 Oct 2026 · News
Anthropic measured its own company and found that agents now log six hours for every hour its people work. The new limit on how much gets done is a person's time to read, check, and say yes.
The most AI-heavy company we can see into has run out of people, not compute. Anthropic's own data says human attention is now the bottleneck, and the company is hiring for judgment while hiring fewer of the juniors who would normally grow into it. That second part is the story.
On Friday Anthropic published How employment is changing at Anthropic, a study of its own staff built from internal hiring records, agent logs, surveys, calendars, and 17 interviews. Agent work hours passed human work hours in March. By August, counting only agents that employees direct themselves, agents logged six hours for every one of their humans. (Anthropic counts agent runtime and skips any wait over 30 minutes; humans are counted at 40 hours a week.)
Then look at what the agents spend their time doing. In interactive sessions, about 80% of the clock is the agent sitting idle, waiting for a person. Anthropic estimated how long it would take to simply read everything its agents write in those sessions, at 250 words a minute. Even for non-technical teams it comes to more than half the work week. One engineer described it as "an army of Claudes, with every one asking for my attention." Another compared it to babysitting, where the kid who wants a juice box opened sounds exactly as urgent as the one who is bleeding.
People are adapting by managing instead of making. Work about the agents (writing specs for them, setting one agent to review another, building dashboards to watch them) went from 10% of the average employee's sessions last fall to 22% this spring. Written standing instructions for agents grew from a few hundred in July 2025 to 8,500 a year later. Meetings are up too: the median employee spends 20% more time in them than a year ago.
Here is the part that should get more attention than the 6 to 1 chart. Junior hires fell from 27% of technical hiring to 9%, and the junior share of technical staff slid from 21% to 14% since early 2025. Hiring managers now write more about a candidate's autonomy and judgment and less about domain expertise. The report is candid that its fixes (stock ownership, an early warning system for at-risk jobs, internal transfers) do nothing for people who never get hired. It also says that in some roles, staff are "increasingly removed from the object-level work that originally built their skills."
Put those lines next to each other. Judgment is the scarce skill. Judgment is built by doing the object-level work. The object-level work is what the agents now do, and the entry-level jobs where people used to do it are the ones being cut. A company can run on that for a while by spending the judgment it already has. It cannot run on it forever.
Two days ago we wrote that AI had made finding bugs cheap and landing the fixes expensive, and closed on "speed upstream turns into judgment work downstream." That was a bank hiring contractors. Now the lab with the most agents has measured the same thing inside its own walls and put a number on it. It also fits the agent that won't take no: an agent pushing past a missing yes is more likely when the person who should say yes has an army asking at once.
The rest of the market is rationing from the other end. Accenture told many staff it will centrally fund only $29 a month of Claude per person, with heavier use billed to a project, 404 Media reported this week. Some had been spending $1,000 a month. Accenture told The Telegraph the figure is one part of its funding approach, not a firm limit. Either way, the meter we wrote about last week has moved from the chatbot plan to the employee. At Anthropic tokens are cheap and people are scarce. At Accenture tokens cost money and the person has to justify them. Both end up asking the same question: whose time is worth the output?
The usual caveat applies, and the report states it plainly: Anthropic is an unusual company, it grew fast, and adopting agents is not random. Fair. But that makes it a leading indicator, not an exception. Most companies will get the agent fleet before they get the fix for who checks its work.
Our call: by the end of June 2027, at least one frontier lab or big tech company launches a named entry-level program pitched as training people to supervise agents, because the judgment pipeline will run dry faster than the agents improve. And the productivity number companies start bragging about will shift from agent hours to human review hours saved. The six-to-one chart is impressive. The number that decides what ships is the other side of the ratio: one person, 40 hours a week.