How work is changing at Anthropic: AI agents working six times more hours than people

The article, published by Anthropic in its economic research section on October 16, 2026, examines how AI has reshaped work and employment within the company itself. The authors argue that because Anthropic sits at the frontier of AI adoption, it can serve as a leading indicator of what will happen at other companies. Note that the text received is cut off mid-sentence near the end, so this summary covers only what appears in the available material.

The study draws on the company's internal data: employment, AI usage, output metrics, employee surveys, internal communication patterns, calendars and hiring-decision records. This data is supplemented by 17 employee interviews. It is an analysis of a single company, one that also builds the technology it studies, and the authors themselves acknowledge methodological limitations, such as the fact that tool adoption is not random.

An initial finding is that Anthropic has hired a smaller share of junior technical roles. According to the article, the share of junior employees in technical departments fell from 21% to 14% since early 2025, and the share of junior technical hires dropped from 27% to 9%. Job openings for junior profiles declined especially sharply from December 2025. Internal discussion about lower demand for junior technical profiles emerged in January 2026. By April, engineering and research teams had stopped prioritizing hires at the lower bands. In a survey that month, a majority of managers said they were looking for a different set of skills because of AI. Around a third of machine learning managers and a fifth of software engineering managers said they were hiring less for that reason.

The weight of technical departments in total headcount also fell, from 61% to 56% over the past year. According to the text, technical hiring is still growing in absolute terms, but hiring in sales, finance and recruiting has grown faster. The authors attribute part of this to the natural expansion of business functions and part to coding tools enabling non-technical employees to build software. Code-focused agent sessions by non-technical employees rose from 2 a month in August 2025 to 90 in May 2026. One salesperson, for example, said he built his own dashboard and workflow in three or four hours.

Within technical teams, the share of employees who do not actively code (no commits in the previous 30 days) fell by almost half in the year to April 2026. The ratio of engineers to product managers is around 17 to 1 and has risen since 2025. Commits by managers have doubled. By contrast, administrative functions that typically grow with a company's size have not expanded at the same pace, thanks to assistant agents, automated review and outsourced tickets. According to the text, Anthropic remains a company with little administrative staff despite its rapid growth.

There is also more internal mobility. Employees with at least a year's tenure were 50% more likely each month to switch departments in 2026 than in 2025. The article describes the case of a product design team that wrote copy for Claude.ai, such as the interface greetings. It developed a context graph that guided agents in writing that content, and then moved on to writing prompts and evaluations for Claude.

On productivity, telemetry data indicate that in March 2026 agent working hours, autonomous and interactive, exceeded human hours for the first time. Looking only at agents directed by individual employees, in August agents worked six times as many hours as their human supervisors. Human hours are calculated from headcount and 40-hour weeks, and agent time counts active execution, excluding waits longer than 30 minutes. Most tokens are spent producing code, but there are also personal agents that manage calendars and messages and prepare meeting summaries.

Code commits per engineer are 5 times faster than six months ago. Data scientists run more data queries, and output also rose in non-technical functions: lawyers processed more contracts, accounts payable handled more invoices and salespeople closed more deals. Because the company is experiencing explosive business growth, the authors try to isolate the effect of AI by comparing 192 engineers who adopted coding agents with 74 who did not during the period, using person and month fixed effects. Code added rises sharply after adoption. To assess quality, they compare code written by agents with code written directly by the same person, in the same month and directory. They find no differences in the share of lines later deleted or reverted, although they caution that quality has many dimensions and that AI errors may be harder to detect. The authors also note that those who adopt fully autonomous agents were already on an upward trajectory before doing so.

In sales, after adopting agents, representatives close deals with more customers, although the authors acknowledge there could be selection bias. One account executive said automations replaced around 20% of his tasks, and another said they allowed him to talk to 30% more customers. Also described are a lawyer who automated the triage of legal queries and a recruiter who used Claude for an analysis that would have taken a couple of weeks. Descriptively, junior employment grows more slowly in roles that use Claude for the most complex tasks.

The text devotes a section to bottlenecks. In an internal survey, 44% of employees agreed in 2024 that processes made them as productive as possible, and 58% in November 2025. In April 2026 the figure rose to 75%, but the question had been softened by removing the "as possible". In August, employees expected AI to be able to do 75% of their work by 2030, compared with 28% of other Americans. When a new model comes out, employees tackle harder problems: the share of tasks above the 90th percentile of the previous model's difficulty rises by around 30%, from 10% to 13%.

Human attention is the bottleneck. Around 80% of the time in interactive sessions, agents are waiting for a human response. One engineer compared the situation to looking after children who all demand attention at once. In the April survey, 27% described productivity gains and 20% said their role had shifted from doing the work to managing Claudes. Reading all the output of interactive sessions would take more than half a working week even for non-technical groups, assuming 0.75 words per token and 250 words per minute.

In response, "meta-work" is growing: building tools, dashboards and oversight for agents. Standing instructions written for agents rose from a few hundred in July 2025 to 8,500 a year later. Meta-work went from 10% of the average employee's interactive sessions in fall 2025 to 22% in spring 2026. One lawyer, for example, began writing frameworks that his clients' agents can discover.

On hiring decisions, managers place more importance on autonomy and judgment than on specific technical expertise and implementation. The article says Anthropic has created an early warning system to monitor roles at risk of disruption and prioritizes internal reassignments as jobs evolve. It also notes that the company generates 2 to 4 times more revenue per employee than Meta and Google did during their rapid-growth phase. The authors conclude that these organizational commitments can shape AI's impact on employment, but raise difficult questions about exactly what they should be. The text breaks off as it begins to address human trust and accountability, which according to the article remain critical despite agents' greater reliability.