Anthropic's interactive sessions wait on people about 80% of the time.

On October 16, Anthropic published a study of how work has changed inside the company now that interactive agent hours outnumber employee hours. The headline numbers are about jobs. The numbers a director should read are about attention: reading all of a non-technical employee’s agent output would take half a work week, and people are spending more time in meetings, not less.

What Anthropic reported

The study draws on Anthropic’s internal records (employment, AI use, output metrics, surveys, internal communication patterns, calendars and hiring decisions) plus 17 employee interviews. All figures below are company-reported:

• Agents outwork people in hours. By August 2026, interactive agent hours outnumbered employee hours 6 to 1.
• Reading the output takes half a week. Even for non-technical employees, reading all the output of their agents would take half their work week.
• The agents mostly wait. Interactive agent sessions spend around 80% of their time idle, waiting for human input.
• Meetings went up. The median employee spends 20% more time in meetings than a year ago. For technical individual contributors, meeting time is up 16% over the past year.
• Managing agents took a larger share of sessions. “Meta-work,” such as writing specs for multi-agent runs and monitoring pipelines, rose from 10% of the average employee’s interactive agent sessions in fall 2025 to 22% in spring 2026. One engineer set up an agent chief of staff to triage agent requests.

The authors say many of Anthropic’s employment practices cannot be replicated across the economy. Treat the numbers as a direction, not a forecast for your department.

Why this lands on a director

The coordination job is moving. It used to be collecting what people did. Now it is deciding which agent output needs a person, and which person.

If nobody makes that call, two things happen. Either someone reads all of it, which on Anthropic’s numbers can take half a work week, or the sorting happens out loud in a meeting. The authors write that “a break from managing the agents typically looks like jumping into a meeting.”

Adding agents without a reading rule adds status meetings.

The read-or-route rule

Before any agent output reaches a person, it gets exactly one of four labels:

1. Decide. It needs a choice only a named person can make. It goes to that person with the options and a due date.
2. Check. It touches money, clients, someone’s pay or job, or anything that leaves the building. A named person reads it before it goes out.
3. Exception. The agent stopped, missed a deadline, hit a limit, or found two sources that disagree. Its owner sees it in the next report.
4. File. Everything else. It is logged and searchable, and nobody reads it by default.

One tiebreaker: if an item cannot be labeled, it goes to the agent’s owner as Decide, and the agent’s instructions get fixed so the next one can be labeled.

A small worked example

This is an illustration, not a client case. Say your department runs three agent routines: a weekly forecast check, a vendor-invoice match and a project-status rollup. Together they produce 60 items a week.

Labeled, that might come out as 4 Decide, 9 Check, 3 Exception and 44 File. Your team now reads 16 items, not 60, and each of the 16 has a name on it.

The Monday status meeting becomes a written report with four short sections:

• Decisions needed: what, who decides, by when.
• Checked and sent: what went out and who read it first.
• Exceptions: what stopped, and who is on it.
• Filed: a count and a link.

If that report is out before 9 a.m., most of the meeting has nothing left to cover.

A test to run this week

Pick one agent routine. Pull one week of its output and label every item with the four labels. Then look at two numbers.

• Share that needs a person. Add Decide and Check together. Our starting line: if that is more than a quarter of the items, the routine is not ready to run without someone beside it. Narrow what it is allowed to do, then test again.
• Share you could not label. Each one is an instruction problem, not a reading problem. Fix the instructions before you add a second routine.

What to do this week

• Choose one agent routine your team already relies on.
• Label one week of its output: Decide, Check, Exception or File.
• Put a named owner on every Decide and Check item.
• Replace one status meeting with the four-section written report.
• Rewrite the agent’s instructions for anything you could not label.

If you want help setting this rhythm across a whole department, a consult block is five hours at $400 an hour. The hours go to the way the department runs its week, and what leaves with you is a written operating rhythm and a delegation matrix for AI and people.