Skip to content
Technology

Claude Spends Most of Its Time Waiting on You

Anthropic's own data says Claude agents spend most of their work time waiting for a person.

4 min read · 979 words · 5 sources
Dario Amodei, Anthropic's chief executive, meets with Rishi Sunak, Demis Hassabis and Sam Altman in London in May 2023
Dario Amodei, Anthropic's chief executive, meets with Rishi Sunak, Demis Hassabis and Sam Altman in London in May 2023. Photo: UK Prime Minister / Wikimedia Commons, CC BY 2.0
“Four of Anthropic's own economists timed Claude's agents at work and found them idle 80 percent of the time, stalled on a reply from a person. Does the company built to make work faster still run at the pace of its slowest human?”

Anthropic employs its own economists. Four of them timed how Claude's agents spend a work session, inside the company's own offices. The agents sat idle for most of it, waiting on a person to answer.

The number is 80 percent1. Claude's agents spend four of every five minutes of a session idle, only waiting for a reply. Anthropic published the finding on Oct. 16, in a report titled "How Employment Is Changing at Anthropic."

80 percent of a Claude agent's session time passed idle, waiting on a person

Share of elapsed time in Anthropic's own interactive, multi-turn Claude sessions, Oct. 16, 2026

The magenta arc is the share of session time Claude's agents spent idle, waiting on a reply; the gray arc is the share spent actively working.

The measure covers Anthropic's internal sessions alone, apart from how outside customers use Claude.

Data through Oct. 16, 2026. Sample: Interactive, multi-turn Claude sessions run by Anthropic's own staff.

Source: Anthropic [1]. Chart by Kelly Dennis.

The numbers behind this chart
ItemValue
Idle, waiting on a person80%
Active work20%
Whole100%

Anthropic runs Claude across its own engineering, sales, legal and research teams. A result this stark, measured by the company that built the product, is hard for a rival lab to dismiss.

Four Economists Timed Their Own Company

Nathan Wilmers teaches work and employment at MIT Sloan1. He wrote the report with three Anthropic colleagues: Regina Catipon, Maxim Massenkoff and Peter McCrory. Their data draws on Anthropic's surveys, its calendars, its hiring records and 17 employee interviews.

The MIT Sloan School of Management building in Cambridge, Massachusetts
MIT Sloan's Building E62, home to Nathan Wilmers, the professor of work and employment who co-wrote Anthropic's report with three of the company's own economists.Kenneth C. Zirkel / Wikimedia Commons, CC BY 4.0

The team studied interactive sessions, the kind where a person checks on an agent mid-task. They measured each agent's active working time. Then they subtracted it from the session's full length1.

What was left over was waiting. "In interactive, multi-turn sessions, the vast majority of agents' session time is not spent working at all, but spent waiting," the report says. One Anthropic engineer put it in three words. People are "bottlenecking the agents," the engineer said.

One engineering group got closer tracking: 192 engineers who adopted coding agents, against 74 who held off1. Both groups answered the same employee survey and sat for the same round of interviews.

The Job Becomes Managing Claude

Anthropic's staff now describe a new kind of workday. Agent work hours passed human work hours inside the company in March2. By August, agents logged six hours of work for every one hour a person logged.

That shift changed what the humans do all day. The share of interactive sessions spent on "meta-work," managing an agent instead of doing the task, rose to 22 percent from 10 percent. A survey in April asked Anthropic staff about their jobs. One in five said their job had turned into managing Claudes.

Dario Amodei speaks on stage at TechCrunch Disrupt in 2023
Anthropic chief executive Dario Amodei on stage at TechCrunch Disrupt in 2023. His company's agents now log six hours of work for every one hour its staff put in.TechCrunch / Wikimedia Commons, CC BY 2.0

Meetings grew too. The median employee spent 20 percent more time in meetings than a year before3. Technical staff logged a smaller rise, 16 percent. Wilmers and his co-authors measured a separate shift inside the company: "employees working at the company for at least a year were 50% more likely to move across departments each month in 2026 than they were in 2025," they wrote1.

The junior share of Anthropic's technical hires fell to 9 percent from 27 percent since early 20253.

Junior hiring falls 18 points while time spent managing agents rises 12

Junior share of Anthropic's technical hires, and share of sessions spent on "meta-work," early 2025 against 2026

Each rod runs from the earlier reading to the 2026 reading; the blue dot is early 2025 and the magenta dot is 2026.

Both rows come from one company's internal records alone, apart from any cross-industry sample.

Data through Oct. 16, 2026. Sample: Anthropic's own technical hiring records and interactive Claude session logs.

Sources: Anthropic [1]; TechCrunch [3]. Chart by Kelly Dennis.

The numbers behind this chart
ItemEarly 20252026
Junior share of technical hires27%9%
Sessions spent managing agents10%22%

Who Is the Real Bottleneck

Anthropic's own reading points at the humans. Wilmers and his co-authors measured people reshuffling roles faster than agents took over those roles outright. Read plainly, the company moves its own staff around faster than it replaces them.

Lorin Hochstein reads the same report and finds a different culprit. He is a staff software engineer at Airbnb, and writes about complex systems on his blog, Surfing Complexity4.

An Airbnb office in Toronto
An Airbnb office. Lorin Hochstein, a staff software engineer at Airbnb, argues Anthropic's agents interrupt at the wrong moments, independent of how fast people answer.Raysonho @ Open Grid Scheduler / Scalable Grid Engine / Wikimedia Commons, CC0

Hochstein argues the delay sits with the agents, independent of how fast a person answers4. "Because the agents don't have a good model of the engineer's current goals and state, they aren't able to make decisions about when it's appropriate to vie for the engineer's attention," he wrote. An agent that interrupts at the wrong moment creates its own delay, whatever a person's actual speed happens to be.

He also faults the report for skipping a harder question4. "Notably, there's no discussion in the paper about whether agent utilization is genuinely important to the business or not," Hochstein wrote. Five times more commits per engineer is a different claim from five times more revenue.

Readers on Hacker News split along the same line. The report's discussion there drew 152 points and 251 comments within a day of posting5. Commenters argued over whether more agent hours signal more output, or just more activity on the clock.

What to do

Ask any vendor selling agents by the hour what share of that time is active work. Anthropic's own 80 percent gives you a number to compare against, from the company that built the product.

Check your team's calendar before you add agents to it. Managing Claude already takes a fifth of a session at the company that makes Claude. Plan for that cost at your company too.

Watch where your own staff move once agents join a team. Anthropic's tenured staff are switching departments half again as often as before. That kind of reshuffling is its own cost, separate from any agent's output, and worth tracking on its own.

By the numbers

  • 80 percent of Claude agent session time passed idle, waiting on a person, inside Anthropic's own interactive sessions1.
  • A ratio of 6 to 1: AI agent work hours over human work hours at Anthropic by August 20262.
  • Sessions spent on "meta-work," managing agents, rose to 22 percent from 10 percent2.
  • Tenured staff were 50 percent more likely to switch departments each month in 2026 than in 20251.
  • Junior hires fell to 9 percent of Anthropic's technical roles, down from 27 percent since early 20253.
  • Meeting time rose 20 percent for the median employee, and 16 percent for technical staff3.
  • 152 points and 251 comments landed on the report's Hacker News discussion within a day5.

What to watch

A second AI lab's own version of this study would show whether 80 percent is normal, or just how Anthropic works. Next year's repeat of Anthropic's own measurement would show whether the wait is shrinking. A rival lab's agent-hour count would settle whether Anthropic stands alone on this number today. Every other lab's own figure is still missing from the public record.

Sources

  1. Nathan Wilmers, Regina Catipon, Maxim Massenkoff, Peter McCrory, "How employment is changing at Anthropic," Anthropic, Oct. 16, 2026, https://www.anthropic.com/research/employment-at-anthropic
  2. Marco Quiroz-Gutierrez, "'An army of Claudes, with every one asking for my attention': Anthropic calls human attention the bottleneck in the botsitting world to come," Fortune, Oct. 19, 2026, https://fortune.com/2026/10/19/work-botsitting-meetings-army-of-claudes-anthropic/
  3. Anthony Ha, "Anthropic says its workforce has become less junior and is spending more time in meetings," TechCrunch, Oct. 17, 2026, https://techcrunch.com/2026/10/17/anthropic-says-its-workforce-has-become-less-junior-and-is-spending-more-time-in-meetings/
  4. Lorin Hochstein, "Anthropic and the changing nature of work," Surfing Complexity, Oct. 18, 2026, https://surfingcomplexity.blog/2026/10/18/anthropic-and-the-changing-nature-of-work/
  5. theanonymousone, "How employment is changing at Anthropic," Hacker News, Oct. 19, 2026, https://news.ycombinator.com/item?id=50139873

Cite this piece

Kelly Dennis, "Claude Spends Most of Its Time Waiting on You," AI Lately, Oct 20, 2026, https://ailately.com/articles/claude-agents-wait-on-humans

Tags: Anthropic · Claude · agent infrastructure · idle time · workplace automation · junior hiring

Related, lately

More

Keyboard

j k
Move through a list
Enter
Open the selected piece
/
Search, and browse the columns
g then r
Right Now
g then a
Articles
g then s
The Signal
g then o
Opinion
g then b
Analysis
g then p
People
?
This sheet
Esc
Close