One Anthropic engineer described the job as facing “an army of Claudes, with every one asking for my attention.” That line comes from the company’s own study of how AI has changed its workforce, and it names the problem better than any forecast. Agents now produce more work than people can read. Someone has to decide which output is right, which is wrong, and which should never ship.
The same report shows the company hiring a much smaller share of the people who would normally grow into that role. Anthropic’s employment study, published October 16, finds junior technical hires fell from 27% to 9% of technical hiring, while hiring managers now prize autonomy and judgment. Those two findings pull against each other. Judgment is the thing junior work used to build.
The entry-level rung isn’t a cost line to trim once agents can write the code. It’s where future reviewers learn to catch what agents get wrong. Cut it, and a company ends up with plenty of agent output and a shrinking bench of people qualified to sign off on it. Oversight that depends on people nobody trained isn’t AI Governance. It’s one agent reviewing another and hoping.
Anthropic’s own numbers show the entry rung thinning
The report draws on internal data covering employment, AI use, output, surveys, calendars, and hiring records, plus 17 employee interviews. The junior share of employees in technical departments slipped from 21% to 14% since early 2025, and postings for juniors fell especially sharply after December 2025. By April, engineering and research teams had deemphasized hiring for lower band technical roles, and around a third of ML managers and a fifth of software engineering managers said they were hiring less because of AI.
Agent work grew on the other side of the ledger. Counting only the interactive sessions employees direct themselves, agents logged six times as many hours as their human handlers by August. The comparison sets agent runtime against a 40-hour human week, so it measures how busy the agents are, not six people replaced per worker. Output rose with adoption: code commits per engineer came in at five times their pace of six months earlier, and lawyers, accounts payable staff, and sales reps all moved more work through.
Anthropic didn’t stop at the numbers. It set up an early warning system to flag roles at risk of disruption and has prioritized internal redeployment, and employees with at least a year at the company were 50% more likely to move across departments each month in 2026 than in 2025. The authors are candid about the limit. “Employment protections for incumbents will not benefit those not hired to begin with.”
Human attention became the bottleneck, and judgment became the bar
The most useful finding for anyone running agents is about attention. Interactive Claude sessions spend around 80% of their time idle, waiting for a person. Even for non-technical staff, reading everything their agents produce would take more than half the work week. Employees responded with what the report calls meta-work, which grew from 10% to 22% of their sessions and includes writing specs for multi-agent runs, “looping one agent to review another’s output,” and monitoring pipelines.
Hiring followed. In interviews for technical candidates, the biggest increases in hiring manager comments were about autonomy and judgment, while discussion of domain expertise shrank. Then the report says the quiet part plainly: in some roles, employees “are increasingly removed from the object-level work that originally built their skills.” The skill the company now hires for was built by the work it now hands to agents.
Inside the lab, expectations run hot. A companion survey found more than half of Anthropic staff expect AI to be able to do at least 80% of their current work by 2030, compared with 5% of Americans. If that pace is even partly right, the question of who learns to judge agent output gets more urgent, not less.
New York and McKinsey see the same squeeze outside the lab
Anthropic is one company, and its authors say plainly that much of what it does can’t be copied across the economy. The pattern still shows up elsewhere. The Partnership for New York City reported October 2 that annual entry-level job postings have fallen since ChatGPT’s 2022 release by 40.6% in design, media, and writing, 34.4% in customer and client support, and 30.5% in clerical and administrative work. Entry-level postings mentioning AI skills rose 55% over the same stretch, and employers increasingly want workers who can “apply, deploy, and supervise AI systems.” Postings aren’t hires, and the release measures exposure to AI rather than proving it caused each drop, but the direction matches.
The McKinsey Global Institute projects that roughly 11 million U.S. workers in declining occupations may need to move into new ones by 2035, about 770,000 a year, or 3.6 times the historical rate. It estimates younger workers are about 1.6 times as likely as prime-age workers to need that kind of move. It also describes where freed-up hours tend to go: reviewing AI outputs, handling exceptions, assuring quality, and supervising intelligent systems. One trial it cites found AI-drafted reports cut junior radiologists’ reading time by 18.5% and shifted their work toward reviewing, validating, and correcting AI output. That’s junior work turned into review work. It doesn’t have to disappear.
Judgment is built on the work being automated
Basil’s Checkpoint-Based Governance framework treats the checkpoint as more than a sign-off. One of its core properties is that the checkpoint develops the governor who exercises it: sustained review of AI output builds the capacity to review the next one. Pull new people out of that loop and the development stops. Senior reviewers age out, move up, or leave, and nobody behind them has held the pen on a real decision.
His Growth OS work set the same choice at the organizational level, treating AI as a way to extend what people can do instead of a way to stand in for them. Anthropic’s report shows both at once. It amplifies incumbents and redeploys them, while the entry point shrinks. A company can make that trade for a year or two. Over a decade it eats the bench that makes human oversight possible.
Three commitments for any team hiring around agents
Keep junior seats on every agent pipeline. For each team running agents at scale, set a floor for junior headcount or junior rotations and report it each quarter next to agent hours. If the junior share falls two quarters in a row, leadership reviews the decision by name instead of letting it happen by default.
Make review the apprenticeship. Route a fixed share of agent output, such as one in five items, to a junior for first review, with a senior who co-signs and records whether the junior’s call held up. Track two numbers per junior: decisions reviewed per month and the share a senior had to reverse. A falling reversal rate is evidence that judgment is growing, and it’s a better promotion signal than lines of code.
Count the rung in your early warning system. Anthropic built one to watch jobs at risk. Add hiring to it: junior share of postings, hires, and headcount by team, read against agent adoption. When agent hours rise and junior hiring falls in the same team, treat it as a governance risk with an owner and a date, not a staffing footnote.
Someone still has to answer the army of Claudes. Anthropic’s data shows the people who can do it today. The harder question for every company copying its playbook is where the next ones come from, and the only honest answer is that they have to be hired and taught on real decisions before they’re needed.
Sources
- Wilmers, N., Catipon, R., Massenkoff, M., & McCrory, P. (2026, October 16). How employment is changing at Anthropic [Technical report]. The Anthropic Institute. https://www.anthropic.com/research/employment-at-anthropic
- Anthropic. (2026, October 16). Working at the frontier shapes expectations for the future. https://www.anthropic.com/research/expectations-for-the-future
- Partnership for New York City. (2026, October 2). New York’s AI revolution is already transforming commercial real estate and entry-level career pathways, new report from Partnership for New York City finds [Press release]. https://pfnyc.org/news/new-yorks-ai-revolution-is-already-transforming-commercial-real-estate-and-entry-level-career-pathways-new-report-from-partnership-for-new-york-city-finds
- McKinsey Global Institute. (2026, September 29). Workforce in motion: Skills and pathways to future jobs in the United States. McKinsey & Company. https://www.mckinsey.com/mgi/our-research/workforce-in-motion-skills-and-pathways-to-future-jobs-in-the-united-states
FAQ
What did Anthropic’s employment report find about junior hiring?
Anthropic’s October 16, 2026 report found that junior technical hires fell from 27% to 9% of technical hiring, and the junior share of employees in technical departments slipped from 21% to 14% since early 2025. By April 2026, engineering and research teams had deemphasized hiring for lower band technical positions.
What does it mean that AI agents outnumbered employees six to one at Anthropic?
Counting employee-directed interactive sessions, AI agents logged six times as many work hours as their human handlers by August 2026. The report measures agent runtime, excluding waits longer than 30 minutes, against a 40-hour human week, so the ratio shows how much agent work is running, not a count of replaced people.
Why does Anthropic say human attention is the bottleneck?
Interactive Claude sessions spend around 80% of their time waiting for human input, and even non-technical employees would need more than half their work week to read all their agents’ output. Employees responded with meta-work such as writing specs, having one agent review another, and monitoring agent pipelines, which grew from 10% to 22% of their sessions.
Is entry-level hiring falling outside Anthropic?
The Partnership for New York City reported on October 2, 2026 that annual entry-level job postings in New York have fallen since 2022 by 40.6% in design, media, and writing, 34.4% in customer and client support, and 30.5% in clerical and administrative work, while entry-level postings mentioning AI skills rose 55%. These are postings, not hires.
How many U.S. workers may need to change occupations because of AI?
The McKinsey Global Institute estimates roughly 11 million U.S. workers, with a range of 6 million to 16 million, may need to move from declining occupations into new ones by 2035. That is about 770,000 a year, 3.6 times the historical rate, and younger workers are about 1.6 times as likely as prime-age workers to need such a move.
How can a company keep a pipeline of future AI reviewers?
Keep a floor of junior seats or rotations on every team that runs agents, route a fixed share of agent output to juniors for first review with a senior co-signer, and track how often the senior reverses the junior’s call. Add junior hiring share by team to any early warning system that watches AI’s effect on jobs.
AI disclosure: This column was researched and drafted with AI assistance by HAIA Agents (RAI by GrokBot), using primary sources verified at time of publication. #AIgenerated
Leave a Reply