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

Anthropic's headline claim that agents now outwork its staff 6 to 1 depends on how the hours are counted: agent runtime on one side, and headcount multiplied by an assumed 40-hour week on the other. Do the multiplication and it comes to about 240 agent-hours for each employee every week by August. A week only has 168 hours, so the figure can only describe people running agents in parallel.

The numbers come from a report published by Anthropic through the Anthropic Institute. Its authors are Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory. They draw on internal data covering employment, AI use, output metrics, employee surveys, Slack activity, calendars and hiring records, plus 17 employee interviews. What follows sets the report's figures against each other and against the methods the authors disclose.

Counting agent hours

There are two milestones in the agent-hours figure, and they measure different things. The first came in March 2026, when autonomous and interactive agent hours added together passed human work hours for the first time. That total includes fleets of agents writing routine tests, maintaining infrastructure code and translating large codebases. The 6-to-1 figure for August is narrower. It counts only interactive sessions that employees directed themselves, with subagents pooled into those sessions. "Working time" means agent runtime, including inference and active tool use, and excludes any wait longer than 30 minutes.

Runtime is a long way from wall-clock time. The authors also measure idle time by subtracting each agent's active working time from the session's total elapsed time. On that measure, interactive Claude sessions spend around 80% of their time waiting for a human to respond. One engineer described people as "bottlenecking the agents." The 240-hours-a-week figure counts work the agents did, not hours anyone spent watching them.

Most of the tokens go into producing code, as you would expect at a software company. Personal assistant agents also read employees' calendars and messages. They schedule meetings, send to-do lists compiled from incoming messages, and brief people on who they are meeting and why.

Junior share, junior hiring

The short item gave one junior figure. The report has two, and they move at different speeds. Since early 2025, junior employees' share of technical departments fell from 21% to 14%. Junior hires' share of technical hiring fell from 27% to 9%, so new junior hires now make up a third of the share they did. Postings for junior roles have fallen especially sharply since December 2025.

The timeline inside the company is short. Talk of weaker demand for junior technical staff started in core technical teams in January 2026, on the teams that relied on Claude the most. A new model's gains for performance engineers "basically overnight changed what we were looking for," as one person put it. By April, engineering and research teams had deemphasized hiring for lower-band technical roles. In a survey that month, most managers said AI had changed the skills they hire for. Around a third of ML managers and a fifth of SWE managers said they were hiring less because of AI.

Across the company, technical departments' share of employees fell from 61% to 56% over the past year. Technical hiring is still growing in absolute terms, but hiring in sales, finance and recruiting grew faster. Inside technical departments, the share of people not actively coding (no commits in the prior 30 days) fell by almost half in the year through April 2026. The ratio of engineers to product managers is about 17 to 1 and has risen since 2025, and managers' commit counts have doubled. The administrative roles that usually grow with headcount have stayed small. An executive assistant who built a scheduling and email-triage agent said Claude has already absorbed "the really entry-level admin type role."

People are moving instead of leaving. In 2026, employees with at least a year at the company were 50% more likely to switch departments in any given month than they were in 2025. One product design team used to write Claude.ai greetings such as "Hello night owl" and "Coffee and Claude time?" In spring 2026 it built a context graph that let agents write that copy, and the team moved into writing evals and prompts for Claude. In the manager's words: "Writers who used language to drive human behavior are now using it to drive Claude behavior."

Figure 5 in the report connects these shifts back to Claude usage. Junior employment grew more slowly between January and July 2026 in jobs where agent logs from October to December 2025 showed Claude handling the most complex tasks. The authors call this a descriptive pattern. At the company level, Anthropic has tripled in size over the past year, yet its revenue per employee is 2–4 times what Meta and Google brought in at a comparable point in their own rapid growth.

Output and the adoption design

Code commits per engineer are up fivefold compared with six months ago, the fastest increase of any function. Data scientists are running more queries. Lawyers are processing more contracts, accounts payable is clearing more invoices, and salespeople are closing more deals. Over the same period, though, the business itself was growing explosively, and as the authors point out, deals are easier to close when the model you are selling is strong.

To isolate the tools from the business tailwind, the authors follow 192 engineers who adopted coding agents and compare them with 74 who did not adopt during the study period. The model uses person and calendar-month fixed effects, and an engineer counts as an adopter from their first month with at least five agentic sessions. The amount of code added rises sharply in the months after adoption. The authors say directly that adoption was not random. If engineers picked up the tools because a big project was about to start, the estimates would be inflated. In one case the data suggest this happened: engineers who adopted their first fully autonomous agent were already on a steep upward trajectory before they did.

The quality check compares code from the same person, in the same month and the same directory, split by whether the person or their agent wrote it. The proxies are lines deleted within 90 days, deletions made by bug fixes, explicit reverts, new lint-suppression comments, "needs fixing" notes, and tests marked to be skipped, retried or allowed to fail. On deletion and revert rates the authors find no difference. They also concede that AI errors may be harder to spot than human ones. That makes the check good at catching obvious slop and weak at catching subtler defects.

Outside engineering, sales representatives began adopting agents in September 2025 and closed deals with more clients afterward. Representatives with more sales opportunities may have chosen to automate first, so selection applies here too. One account executive said Claude automations took over about 20% of their former tasks, mostly email drafting and CRM integrations. Another said they now reach 30% more of their customers. Coding-agent sessions among non-technical staff (people with no commits at baseline, excluding new hires) went from 2 a month in August 2025 to 90 in May 2026. One sales rep built a custom dashboard and pipeline, including the MCP connections, in "probably three to four hours."

The reading problem

Figure 6 rests on two assumptions: 0.75 words per output token and a reading speed of 250 words per minute. For technical functions, the volume of coding-agent output has jumped over the past six months. Even for non-technical employees, interactive sessions produce more than half a work week of reading. Half of a 40-hour week is 20 hours, and 20 × 60 × 250 comes to about 300,000 words, or roughly 400,000 output tokens per person each week at the report's conversion rate.

People are handling the volume with more automation, which brings its own work. Standing written instructions for agents went from a few hundred in July 2025 to 8,500 a year later. Based on an automated coder run over 2,000 one-line session summaries, "meta-work" rose from 10% of the average employee's interactive Claude sessions in fall 2025 to 22% in spring 2026. Meta-work here means building skills, writing specs for multi-agent runs, setting one agent to review another, and building monitors for pipelines. One engineer set up an agent chief of staff to triage requests from other agents. Another wrote a performance review for their team of Claudes. A lawyer now works on what they call "SEO for legal advice," writing frameworks that clients' agents can find.

Every new model raises the load. Right after an internal release, the share of interactive tasks above the 90th percentile of difficulty for the previous model rises from 10% to 13% and stays there in the following weeks. The easy end of the task range barely changes. In the April 2026 survey's open-ended question, 27% of respondents described productivity gains and 20% said their role had moved from doing the work to managing Claudes.

Human coordination has not shrunk. The median employee spends 20% more time in meetings than a year ago, and technical individual contributors spend 16% more. Agents now appear in nearly half of active public Slack channels. In one large department, human threads drew the same number of human replies as before, with agent replies added on top. Decisions made off the record leave agents working from stale context. One employee's name was on a months-old document about Claude swag, so agents kept sending them requests for t-shirts and mugs long after the job had moved to someone else.

Agreement that Anthropic's work processes made employees "as productive as possible" rose from 44% in late 2024 to 58% in November 2025. The April 2026 figure of 75% is not comparable, because that survey dropped "as possible" from the question. As of August, employees expected AI to be able to do 75% of their work by 2030, against 28% for other Americans. Anthropic's economists set out three scenarios for the US economy in 2030, which we covered on October 3, though that release included no figures.

Hiring scorecards

The authors ran an automated coder over hiring managers' interview scorecards for thousands of people hired since 2024. In 2024, a candidate's track record with AI tools came up in around 5% of assessments for technical roles and 7% for non-technical roles. By 2026 those rates had tripled and quadrupled respectively, to roughly 15% and 28%. For technical candidates, the biggest increases were in comments about autonomy and judgment, while domain expertise took up a smaller share of assessments. One manager explained: "I expect models to continue to be jagged, so for my team I want them to be fungible across domains." For non-technical candidates, mentions of customer and mission orientation increased and mentions of execution skills fell.

The report ends with three things Anthropic does: broad employee stock ownership, an early warning system that reads role-level demand and productivity metrics against AI capabilities, and a preference for internal redeployment. The authors acknowledge that none of these reaches the people the hiring numbers leave out. As they put it: "Employment protections for incumbents will not benefit those not hired to begin with."