Anthropic's AI agents wait on people 80% of the time

The company that makes Claude has measured its own staff, and the slow part of its AI work is now people. In interactive sessions, Anthropic’s agents spend about 80% of their time waiting for a human to respond. If your team is adding agents, that number is the one to plan around. More agents will not help much if nobody has time to read what they produce.

The figure comes from “How employment is changing at Anthropic”, a report published on October 16 by four researchers at the Anthropic Institute. It draws on the company’s internal data on hiring, AI use, output, surveys, calendars and Slack, plus 17 employee interviews. Most coverage, including ANI’s, led with the junior-hiring numbers.

We think the more useful part for a smaller company is how the work itself changed. Below is what happened, why it happened and what to take from it.

Agents now out-work the people who direct them

By August, agents directed by individual employees at Anthropic worked six times as many hours as the people handling them, according to the report. That count covers only employee-directed interactive sessions, and it assumes each person works 40 hours a week.

Output rose too. Code commits per engineer reached five times their pace of six months earlier.

The gains reached sales too. One account executive said Claude automations took over around 20% of their previous tasks, such as drafting emails. Another said similar automations let them talk to 30% more of their customers.

Quality is the obvious worry. The report compared code written by an agent with code written by hand by the same person, in the same month and the same directory. It found no difference in how often the lines were later removed or reverted.

The researchers add a caution: AI errors may be harder to spot than human ones. So “no difference” here means no difference on the proxies they could measure.

The bottleneck moved to reading

Faster agents create a new problem. Somebody has to read and answer them. One engineer described facing “an army of Claudes, with every one asking for my attention.” Another said people are “bottlenecking the agents.”

The report puts a size on it. It converted each session’s output into reading time, at 0.75 words per token and 250 words a minute. Even for non-technical teams, reading everything their interactive sessions produced would take over half the work week.

The load also keeps climbing. After new model releases, staff hand agents harder work: the share of tasks above the old model’s 90th percentile of difficulty rose from 10% to 13%.

So the 80% idle figure is not a sign of lazy software. It is a queue. The agent has finished its part and is waiting for a person to look, decide and send it on.

Managing the agents became a job of its own

People at Anthropic responded by building tools to manage their agents. The report calls it “meta-work”: writing specs for multi-agent runs, setting one agent to review another, and watching pipelines. It grew from 10% of the average employee’s interactive sessions in fall 2025 to 22% in spring 2026.

Written standing instructions for agents grew from a few hundred in July 2025 to 8,500 a year later. One engineer set up an agent “chief of staff” to sort incoming agent requests. Another wrote a performance review for their team of Claudes.

The trade-off is plain. That time comes out of the work itself. A fifth of a person’s agent sessions spent on tooling is time they are not spending on customers or code.

Meetings went up, and decisions had to be written down

You might expect agents to cut meetings. The opposite happened. Based on calendar data, the median Anthropic employee spends 20% more time in meetings than a year ago, and technical individual contributors 16% more. The report says part of that is company growth and a shift toward roles that meet more.

Agents also joined the conversation. They now appear in nearly half of Anthropic’s active public Slack channels. Human threads kept the same number of human replies, with agent replies layered on top.

That creates a new rule of thumb. Agents only know what is written where they can read it. One employee’s name stayed on a months-old document about company swag, so agents kept sending them requests for t-shirts and mugs long after the job had moved on. The report puts it bluntly: private work in unlogged meetings and direct messages “is invisible, confusing and discouraged in an AI-heavy workplace.”

Hiring shifted toward judgment, and away from juniors

The hiring data is the starkest part. The junior share of Anthropic’s technical hires fell from 27% to 9%. The junior share of staff in technical departments fell from 21% to 14% since early 2025. In an April 2026 survey, around a third of machine learning managers and a fifth of software engineering managers said they were hiring less because of AI.

What managers look for changed too. In 2024, interview scorecards mentioned a candidate’s record with AI tools in about 5% of technical assessments and 7% of non-technical ones. By 2026 those rates had tripled and quadrupled. For technical candidates, the biggest increases were comments on autonomy and judgment, while domain expertise shrank as a share.

The authors name the gap in their own response. Anthropic tracks roles at risk and moves people between teams, but “employment protections for incumbents will not benefit those not hired to begin with.”

Read it as one company’s evidence

This is a vendor studying its own use of its own product, and that deserves weight. Anthropic tripled in size in the last year, so some of its output growth may have come from business growth alone. The authors say so, and add that “tool adoption is not random.” Engineers who took up fully autonomous agents were already on a steeply rising trajectory before they did.

The report also says many of Anthropic’s practices, including wide stock ownership, “cannot be replicated across the economy.” A ten-person firm is unlikely to have 8,500 standing instructions or a research team counting idle minutes.

What carries over is the mechanism. Agents produce faster than people can read, and the cost shows up as review time, tooling time and meetings.

What this means for you

The decision most teams face is whether to add more agents or make better use of the ones they have. Anthropic’s data says the limit will be your people’s attention, so plan for that first.

• Estimate the reading load before you scale. Use the report’s own conversion: output tokens times 0.75 gives words, and 250 words a minute gives reading time. By that math, one million output tokens is 750,000 words, or 50 hours of reading.
• Budget time for managing agents. At Anthropic it reached 22% of sessions. If nobody’s week has room for writing instructions and checking pipelines, the agents will mostly wait.
• Write decisions down where agents can read them. Keep meeting notes and ownership documents current. A stale owner on a document will keep drawing requests, as the swag example shows.
• Decide how you will train juniors. If agents take the entry-level tasks, write down which work new hires will still learn on before you cut those roles. The report’s authors admit their own measures do not solve this.

The trade-off cuts both ways. More review keeps quality up but caps how much the agents deliver. Less review frees people but lets more errors through, and the report warns those errors may be harder to see.

At Anthropic, the agents are already waiting on people 80% of the time. That is the number to check in your own team before you add the next one.

Sources

• Anthropic Institute: How employment is changing at Anthropic (Wilmers, Catipon, Massenkoff, McCrory, October 16, 2026)
• Open, with ANI: Junior tech hires’ share at Anthropic falls from 27% to 9% as AI reshapes work