An AI lab published what its own agents did to its workforce. The numbers are a warning for early-career workers and a planning tool for anyone who hires.

The Short Version

On Friday, Anthropic published a study of its own employees, and one number jumps out. Among the people the company hired into technical roles, the share at the junior level fell from 27% to 9%. Postings for junior roles dropped sharply after December 2025, and about a third of its machine-learning managers told an internal survey in April that they were hiring less because of AI.

This is one company, and an unusual one. It builds the tools it measures, it tripled in size over the past year, and its own researchers warn it may represent no one else. Still, it's the most detailed look yet at a workplace where software agents now log more hours than people. The report also shows where the work went: senior staff spend more of their week reviewing, specifying and waiting on agents, and fewer of the starter tasks are left for beginners. An IMF report released the same week in Bangkok found a smaller, steadier effect across Asia. Read together, the two give early-career workers and the people who hire them something concrete to watch.

What the Report Measured

The study, titled "How employment is changing at Anthropic," comes from four of the company's researchers. It draws on headcount records, agent usage logs, code output, employee surveys, calendars and thousands of hiring scorecards, plus 17 interviews. Anthropic posted it Friday and updated it Saturday.

Its biggest finding is about time. Across the company, hours of agent work passed hours of human work in March 2026. By August, in sessions employees ran themselves, agents logged about six times the hours of the people directing them. That comparison counts each employee as a flat 40-hour week, so it's a rough yardstick, though the direction is hard to miss. Code commits per engineer ran about five times their level of six months earlier.

Output and value aren't the same thing, and the authors say so plainly. Engineers who adopted agents early were already on a steeper path, the company's fast growth muddies any trend, and more code may or may not mean better code or more business impact. They call the junior-hiring analysis descriptive, so it shows a pattern without proving AI caused it. On drawing lessons for the wider economy, their advice is to "proceed carefully."

Where the Junior Jobs Went

The junior figures reward a slow read. The share of junior-level people across technical departments slipped from 21% to 14% since early 2025. The share of new technical hires who were junior fell faster, from 27% to 9%. Both are shares. Technical hiring at Anthropic grew in absolute terms while the company tripled, and the report gives no count of junior hires, so the raw number may have fallen less than the percentage suggests. The report also leaves "junior" undefined. What it does show clearly is the mix: each new technical team gets built with far fewer beginners on it.

The reason shows up in how experienced people now spend the day. Interactive sessions with the company's AI sit idle about 80% of the time, waiting for a human to respond. One engineer described the job as "an army of Claudes, with every one asking for my attention." Building tools, specs and monitors for agents rose from about 10% of an average employee's interactive sessions last fall to 22% this spring, and median meeting time climbed 20% in a year. Agents now absorb the tasks a new hire used to learn on, and the scarce work is judgment, review and coordination.

Hiring changed to match. Mentions of a candidate's use of AI tools in interview scorecards roughly tripled for technical roles and quadrupled for other roles between 2024 and 2026, from a starting point of about 5% and 7%. Technical assessments now weigh autonomy and judgment more heavily than deep expertise in one field. A candidate walking into that interview is being tested on whether they can direct an agent and catch its mistakes.

The Wider View From Asia

Anthropic sits at the far edge of adoption, so a broader dataset helps. The IMF's Regional Economic Outlook for Asia and the Pacific, launched at the IMF and World Bank meetings in Bangkok, looked at about 40 million online job postings across 19 Asian economies, according to India's ANI news agency. Since ChatGPT launched in late 2022, employer demand in highly AI-exposed occupations fell nearly 10% compared with less exposed jobs. That decline appears to have stabilized, even as the tools kept improving. Firms that adopted AI also showed stronger hiring growth than firms that held off, including in the exposed jobs.

Krishna Srinivasan, who heads the IMF's Asia and Pacific Department, argued that AI is reshaping work more than eliminating it, China Daily reported. He called tackling youth unemployment and skills mismatches a necessary reform for the region.

The two studies measure different things, one company's hiring mix and a region's job ads, so their numbers can't be compared directly. They still point the same way. Demand shifts first at the entry points of exposed jobs, and the firms that adopt AI keep hiring, for a different blend of skills. For most readers the IMF's 10% is the steadier signal, and Anthropic's 9% is a preview of what an aggressive adopter looks like.

A Hiring Decision, Walked Through

Picture Maria, who runs a 30-person software consultancy and budgeted for two junior developers this winter. Her senior engineers tell her agents can cover the work. Before she cancels the hires, she needs a baseline: the last two quarters of tickets closed per engineer, and how often finished work came back as a bug. She also needs her team's agent use logged well enough to separate agent work from human work.

A bounded test would keep one junior hire and pair that person with agents for 90 days, while a senior engineer handles a comparable workload with agents alone. Maria would count accepted results as code that ships and survives 90 days without being reverted or rewritten, which mirrors the quality checks Anthropic used. She'd also track the cost that tends to hide: senior hours spent reviewing agent output and answering its questions. Anthropic's idle-time and meeting figures suggest that load grows quickly.

The test can fail in two directions. The senior-plus-agents setup can look productive while the review queue turns into a bottleneck and quality slips three months later. Or the junior can look slow in month one and get cut before learning anything. The longer risk sits further out, because a firm that hires no juniors has nobody to promote into senior roles in five years. Maria's engineering lead should hold written stop criteria, such as a revert rate above the old baseline, and the authority to end the test or reopen junior hiring.

Opportunity Radar

Early-career candidates now face interviews that score how they work with AI, and most have no good way to show it. A small service could run realistic, timed work samples where a candidate directs an agent through a task, catches its errors and explains the tradeoffs, then receives a scored portfolio to share. University career centers and coding programs are the likely first buyers, since their placement numbers ride on it. Test it with one career center and 20 students this semester, then ask five hiring managers whether they'd read the results. If the managers shrug, walk away. Demand's just a hypothesis until employers say they'd actually use it.

What You Can Do With This

If you're early in your career

Build proof that you can direct and check AI work: a project where you used an agent, found its mistakes and fixed them. Anthropic's scorecards suggest that skill is now part of the interview. In interviews, ask how juniors get trained alongside agents, and favor employers with a real answer.

If you hire

Measure before you cut. Track review hours and rework next to output, and keep at least a small junior pipeline while you test, because the people you skip hiring now are your senior bench later.

If you're a parent or teacher

Treat the 27%-to-9% figure as one company's data and the IMF's steadier 10% as the wider picture. Both reward judgment, clear communication and checking work, and students can practice those on any assignment that asks them to evaluate an AI's answer.

If you manage a team that uses agents

Watch where your senior people's hours go. If more of the week is review and coordination, build that into workloads and promotion criteria and count it as real work.

The Bigger Picture

Anthropic says it's standing up an early warning system to track jobs facing AI-driven change, it's prioritizing internal moves for current staff, and it wants "farsighted government policy around retraining and income support." The report concedes the company's own steps won't prevent AI from disrupting jobs. That candor points at a gap. A company can retrain an employee. No retraining program reaches the applicant whose role was never posted. If other employers adopt agents the way Anthropic has, the first visible effect may be a thinner entry rung across many fields, and the new on-ramps will have to be built by schools, employers and governments together.

References

Anthropic, How employment is changing at Anthropic (Wilmers, Catipon, Massenkoff, McCrory), October 16, 2026, updated October 17, 2026 Anthropic, How Employment is Changing at Anthropic, full report PDF, October 2026 ANI, AI reshapes labour demand across Asia, posing challenges for job creation and skills development: Report, October 18, 2026 China Daily, Asia-Pacific urged to ready workers for AI, October 16, 2026

Read on the web: https://theainextwave.com/editions/2026-10-18/