Junior technical hires at Anthropic fell from 27% to 9%

Junior roles at Anthropic fell from 27% to 9% of technical-department hires since early 2025, and over the same stretch the agents its employees run came to log six times their human hours, according to an October 17 report by Anthropic Institute researchers published by Anthropic. Both figures come from internal company data. Still, the 9% is a share of a hiring pool that kept growing, and the productivity gains come from comparing engineers who adopted agents with those who hadn’t, groups the authors say were not formed at random.
What the 6-to-1 and the 9% measure
The agent figure covers only interactive sessions that employees directed, with subagents pooled together. Agent time is runtime, inference plus active tool use, minus any wait longer than 30 minutes. Human time is headcount multiplied by an assumed 40-hour week. If you add the fleets of autonomous agents that write tests, maintain infrastructure code and translate codebases, combined agent hours passed human hours for the first time in March 2026.
The junior numbers come in three layers. Juniors’ share of headcount in technical departments slipped from 21% to 14%, their share of technical hires dropped from 27% to 9%, and junior postings have fallen especially sharply since December 2025. Internal talk of lower demand for junior engineers started in January 2026 on the teams that leaned on Claude hardest. One person put the trigger down to the uplift newer models gave performance engineers: it “basically overnight changed what we were looking for.” When managers were surveyed in April, most said AI had them hiring for different skills, and around a third of ML managers and a fifth of SWE managers said AI had them hiring less.
Anthropic tripled in size over the past year, and technical hiring kept growing in absolute terms. Sales, finance and recruiting grew faster, so technical departments went from 61% to 56% of employees. Part of that shift may come from non-engineers writing software themselves: among employees with no code commits at baseline, coding-focused agent sessions went from 2 a month in August 2025 to 90 in May 2026. The company also takes in 2 to 4 times as much revenue per employee as Meta and Google did at a comparable point in their own growth, the report says.
192 adopters, 74 holdouts
Code commits per engineer are running at 5x the pace of six months ago. Lawyers processed more contracts, accounts payable moved more invoices and salespeople closed more deals. The authors flag the confounder themselves: “It is easier to close deals when you sell a strong model.”
To separate tools from tailwind, the researchers followed 192 engineers through the month they adopted coding agents, counted from their first month with at least five agentic sessions, and set them against 74 engineers who did not adopt during the study period. Both groups worked at the same company in the same months. The code engineers added rose sharply after adoption. The caveat sits in the method: someone about to start a big project may tool up for it, and that would bias the estimate. Engineers who adopted their first fully autonomous agent were already on a steeply upward trajectory compared with non-adopters, which the authors read as those employees already ramping up their output.
On quality, the team compared code from the same person, in the same month and directory, written by an agent versus by hand. They found no difference in how many lines were later removed or reverted, while noting that AI errors may be harder to spot than human ones. Sales got the same before-and-after treatment, with the same selection caveat. One account executive said Claude automations took over around 20% of their former tasks.
An army of Claudes
About 80% of interactive Claude session time is spent idle, waiting for a human. Converting output tokens to words at 0.75 words per token and reading at 250 words a minute, the authors estimate that reading everything their interactive sessions produce would take non-technical staff more than half the work week. Technical functions have seen that volume jump over the past six months. One engineer described facing “an army of Claudes, with every one asking for my attention.”
The response is what the report calls meta-work: writing specs for multi-agent runs, having one agent review another and monitoring pipelines. It rose from 10% of the average employee’s interactive sessions in fall 2025 to 22% in spring 2026, judged by an automated coder on a sample of 2,000 one-line session summaries. Written standing instructions for agents grew from a few hundred in July 2025 to 8,500 a year later. Meetings did not shrink either. The median employee spends 20% more time in them than a year ago, and technical individual contributors 16% more.
Agents also changed what counts as visible work. They showed up in nearly half of active public Slack channels, mostly in dedicated human-Claude channels, and human threads kept the same number of human replies. A decision made in an unlogged meeting never reaches the agents, though. One employee whose name sat on a months-old document about Claude swag kept getting agent requests for t-shirts and mugs long after the job had moved on.
What interviewers write down
Anthropic ran an automated coder over hiring managers’ interview scorecards for thousands of hires since 2024. In 2024, a candidate’s record with AI tools came up in around 5% of assessments for technical roles and 7% for non-technical ones. By 2026 that had tripled for technical roles and quadrupled for non-technical ones. For technical candidates the biggest gains were in comments on autonomy and judgment, while domain expertise shrank as a share of the assessment. “I expect models to continue to be jagged, so for my team I want them to be fungible across domains,” one manager said. For non-technical candidates, customer and mission orientation came up more and execution skills less.
The junior gap
The measures the company describes are aimed at people it already employs. It is building an early warning system that reads role-level demand and productivity metrics against AI capabilities and use, and it prioritizes internal moves. Employees with at least a year at the company were 50% more likely to switch departments each month in 2026 than in 2025. One product design team that used to write Claude.ai greetings now writes evals and prompts for Claude. The authors also credit widespread stock ownership, while acknowledging it is an easy posture for a company whose valuation has risen.
On the hiring decline itself, the authors concede the point directly: “Employment protections for incumbents will not benefit those not hired to begin with.”
Up next
Replit’s CEO says it left Anthropic. Insiders say otherwise
Replit Agent ran Claude Fable 5 in production as of late September, per Replit’s own blog, yet CEO Amjad Masad now says Replit dropped Anthropic for ChatGPT over cost and performance. Which model runs your coding agent right now, Claude or ChatGPT?
Related stories
- Scientists use Claude least where they say it matters most
- Anthropic economists sketch three AI futures for 2030
- Anthropic bought a biotech with one full-time employee
- Anthropic’s IPO filing says a key safety threshold is met
- Anthropic buys Mithril, a one-person biotech run on Claude
- Anthropic’s new wet lab puts Claude in charge of robots
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

