TL;DR: Anthropic published a report on how AI agents changed work inside the company. By August, agents launched by employees logged six times as many hours as the employees did. Junior hires fell from 27% to 9% of technical hiring. Output rose, but the new limit is people: agents spent about 80% of interactive sessions waiting for a human. For a small business, the lesson is to plan for review time and written instructions, not just more agents.

What happened

The report, “How employment is changing at Anthropic,” came out on October 16 (US time). Four Anthropic researchers wrote it using internal data on headcount, AI use, output, surveys, calendars and hiring records, plus 17 employee interviews. Two days later, on October 18 (US time), Anthropic announced its Public Benefit Fund, which names labor-market impacts as one example of what the fund could address.

The main numbers, all from Anthropic’s report:

  • Agents outwork people on paper. In March 2026, total agent work hours passed human work hours for the first time. Counting only sessions that employees ran themselves, agents worked 6 hours for every human hour by August. Anthropic counts agent runtime and assumes 40 hours a week per employee.
  • Fewer junior technical roles. Juniors’ share of technical staff fell from 21% to 14% since early 2025, and their share of technical hires fell from 27% to 9%. Anthropic doesn’t attribute the drop to agents alone: it says the shift happened as agent use grew, and about a third of ML managers and a fifth of software managers said they were hiring less because of AI. Technical teams’ share of all staff fell from 61% to 56% as sales, finance and recruiting grew faster.
  • More output. Code commits per engineer are 5x what they were six months earlier. Lawyers processed more contracts, and salespeople closed more deals after adopting agents. Comparing 192 engineers who adopted coding agents with 74 who didn’t, output rose after adoption. Anthropic found no difference in how often agent-written code was later removed or reverted.
  • Non-coders now build software. Coding-agent sessions by staff who didn’t write code before rose from 2 a month in August 2025 to 90 in May 2026.

Why it matters

The biggest finding for anyone running agents is the bottleneck. Anthropic says agents spent around 80% of interactive session time idle, waiting for human input. Even for non-technical teams, reading everything their agents produced would take more than half the work week.

Staff responded with what the report calls “meta-work”: writing instructions, setting one agent to check another, and building dashboards. That share of sessions rose from 10% 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. Meetings didn’t shrink either: the median employee spends 20% more time in meetings than a year ago.

Hiring changed to match. Hiring managers’ interview notes on candidates’ AI tool use tripled for technical roles and quadrupled for non-technical roles between 2024 and 2026. For technical roles, comments on autonomy and judgment rose the most, while domain expertise got less weight.

Read these numbers with care. Anthropic is an AI lab studying its own staff, the data is internal and not independently audited, and the report itself says adoption wasn’t random and many of its practices “cannot be replicated across the economy.”

What to do

  1. Budget for review, not just agents. If your agents will wait on you most of the time, more agents won’t help. Decide who checks each agent’s output and when. Our guide on stopping agents from trusting wrong tool outputs has a prompt for adding a check step.
  2. Automate the second repeat. One Anthropic salesperson’s rule: “anytime you do anything a second time,” ask how Claude can handle the third. Another account executive said Claude automations took over about 20% of their prior tasks.
  3. Write things down where agents can read them. Standing instructions do the heavy lifting. Anthropic also found that decisions made in unlogged meetings leave agents working from old context, so keep owners and decisions current in shared docs.
  4. Let one agent check another. For code, that can be a managed reviewer like Claude Code Review, which costs extra per review.
  5. Hire and train for judgment. If agents handle the routine work, the useful skill becomes knowing what to ask for and when output is wrong.

Sources & method

Sources

Method. Last checked October 19, 2026. All numbers and quotes come from Anthropic’s published report and posts; we haven’t verified Anthropic’s internal data. Dates are US dates as shown on Anthropic’s pages. No affiliate links on this page. How we work.

Spotted an error? See our corrections policy and tell us.