Junior Hires at Anthropic Fell From 27% to 9% as Agents Took the Entry Work

Junior engineers dropped from 27% to 9% of Anthropic's new technical hires — the sharpest workforce signal in a report the company published on its own operations. The methodology: internal headcount data, agent session telemetry, calendar and Slack patterns, hiring scorecards, plus 17 employee interviews. Because Anthropic has rebuilt more workflows around agents (autonomous software that executes multi-step tasks) than almost any B2B company, this is the most granular look yet at what agent adoption does to hiring and output.

How far did junior hiring actually fall?

Since early 2025, juniors fell from 21% to 14% of technical department headcount, and new technical hires who were junior fell from 27% to 9%. Junior postings dropped especially hard after December 2025. The shift began in January 2026 among teams that relied most on Claude, with coding and bug-finding moving to agents first. By April 2026, a majority of managers said they were hiring for different skills because of AI; about a third of ML managers and a fifth of software engineering managers said they were hiring less because of it. Anthropic concedes the point directly: redeployment programs help current payroll, and do nothing for the junior engineer who never got an offer.

Did the company shrink?

No — headcount tripled in a year. At age five, Anthropic runs 2 to 4 times the revenue per employee that Google and Meta did at the same stage, adjusted to 2025 dollars. Technical departments fell from 61% to 56% of headcount while sales, finance, and recruiting grew faster; clerical roles barely grew at all. Kelly Loftus scaled startup sales from under 10 people to 150+ in 18 months and told SaaStr AI 2026: "we still don't really have quotas, we have shadow targets." Eleanor Dorfman rebuilt the sales org in January 2026 because the company couldn't absorb 3 to 5 times the reps fast enough; four months later, 54% of new enterprise logos came in self-serve. The authors flag that 54% as an outlier driven by extreme demand — don't build your plan around it.

Are agents really outworking humans?

By hours, yes. Agent work hours passed human work hours in March 2026; by August, agents worked 6 times the hours of the humans directing them. Code commits per engineer hit 5x their level from six months earlier. In a tracked comparison of 192 engineers adopting coding agents against 74 who hadn't, output rose sharply after adoption — and agent-written code was reverted at the same rate as the engineer's own hand-written code. But interactive sessions sit idle about 80% of the time, waiting for a human. One engineer's phrasing: people are "bottlenecking the agents." Converting agent output into reading time, even non-technical teams would spend more than half the work week just reviewing it. The authors add a causality caveat — early adopters were already on steep output trajectories, so treat the productivity numbers as directional.

What's replacing the old work?

Meta-work — the job of managing agents. Standing written instructions for agents grew from a few hundred in July 2025 to 8,500 a year later, and meta-work rose from 10% to 22% of interactive Claude sessions between fall 2025 and spring 2026. Meetings are up too: 20% more time for the median employee, 16% for technical ICs, because cross-team decisions still happen in a room — and now must be written down where agents can read them. Unrecorded decisions leave agents working from stale context; one employee's name on an old swag document kept triggering agent requests for t-shirts and mugs.

What do interviewers score now?

Anthropic ran its 2024-onward hiring scorecards through an automated coder. Mentions of candidates' AI tool track record tripled for technical roles and quadrupled for non-technical ones between 2024 and 2026. Autonomy and judgment rose most; domain expertise shrank as a share of the assessment. One manager's rationale: models stay uneven across domains, so the team needs people who can move between them.

Smaller signals worth tracking: an AE said automating email and CRM work let them reach 30% more customers; non-technical employees' coding sessions went from 2 to 90 per month in nine months; the engineer-to-PM ratio sits near 17 to 1 and rising; and after each model release, the share of very hard agent tasks rises about 30%, from 10% to 13%, and stays there.

Anthropic pairs all this with broad equity — the authors compare it to the 1950s union bargain of productivity-linked pay — and an early warning system that tracks roles against AI capabilities and moves people before a role disappears. Most companies can't match the equity. Tracking which roles agents are absorbing costs almost nothing, and the next model release will keep raising the share of hard work agents can hold.