Anthropic says juniors’ share of technical hires fell from 27% to 9%

At Anthropic, producing work is increasingly an AI task; deciding what deserves attention remains a human one. In a workplace study published October 16, the company described rising output alongside a smaller share of junior technical hires and more time in meetings—a picture of automation changing people’s jobs rather than simply clearing their calendars.

The report combines internal employment records, AI-use data, output measures, surveys, calendars and hiring decisions with 17 employee interviews. Anthropic presents its heavily automated workplace as a possible early signal of wider labor-market changes. Its findings describe one company that develops the tools it uses, not an economy-wide result.

The work grows faster than the attention

AI agents—software that carries out tasks using tools—now operate at a scale that exceeds employee working time. By August, employee-directed interactive agents logged six times as many hours as their human handlers, Anthropic says. That comparison counts agent runtime, including active tool use, against an assumed 40-hour employee week; it is not a measure of equivalent work.

The resulting volume creates another job: keeping track of it. The report estimates that even non-technical employees would need half their workweek to read everything their Claude agents produce. Employees respond by building visibility tools and monitoring systems for agent workflows, making human attention and judgment a constraint on further output.

In an April survey, 20% of employees said their jobs had shifted from doing work to managing Claudes. Coordination has not disappeared, either. TechCrunch’s account of the study says calendar data showed meeting time per employee rising 20% year over year, and 16% for technical contributors. Anthropic attributes that increase to the continuing importance of human trust, accountability and face-to-face coordination.

Hiring shifts toward judgment

By April, engineering and research teams had reduced their emphasis on lower-level technical hiring. Around a third of machine-learning managers and a fifth of software-engineering managers reported hiring less because of AI. Hiring managers increasingly emphasized autonomy and judgment over specific technical expertise and implementation skills.

That does not mean technical headcount is shrinking. Technical hiring continued growing in absolute terms, but sales, finance and recruiting grew faster. Technical departments’ share of employment fell from 61% to 56%. Anthropic attributes part of that shift to normal business expansion and suggests another part could reflect non-technical staff building their own software.

Existing employees are moving, too. Staff with at least a year at Anthropic were 50% more likely to change departments each month in 2026 than in 2025. The report gives concrete examples of how automation redirects work:

• A product-writing team built a system to guide agents in writing interface content, then moved into prompts and model evaluations.
• An executive assistant automated scheduling and email triage, shifting attention toward higher-level chief-of-staff tasks.

More output, with limits on the conclusion

Anthropic says code commits per engineer reached five times their pace six months earlier. Lawyers processed more contracts and salespeople closed more deals. To separate tool adoption from the company’s business growth, researchers followed 192 engineers adopting coding agents and compared them with 74 who did not adopt during the study.

Code added rose sharply after adoption. But employees chose whether to adopt: someone preparing a large project might also start using agents, biasing the estimate. Anthropic also cautions that more code need not mean better code or greater business impact.

Its quality checks compared human-written and agent-written code from the same employee, month and coding directory. The company found no difference in subsequent removal or reversal rates. Those are useful checks, not a complete verdict: the report acknowledges that quality has many dimensions and AI errors may be harder to identify.