By Tencent Technology
Anthropic on Oct. 16, U.S. time, released an internal report titled How Employment is Changing at Anthropic, laying out how heavy AI adoption has changed job content, hiring patterns, and labor demand inside the company.
The study draws on Anthropic’s internal employment records, AI usage telemetry, employee output metrics, staff surveys, internal communications, calendar data, and hiring scorecards. It also incorporates 17 in-depth employee interviews.
The report presents four main findings. First, the share of entry-level technical hiring fell sharply, from 27% in early 2025 to 9% in 2026, even though Anthropic’s overall headcount continued to expand. Second, AI agent labor rose quickly. By August 2026, interactive agents directly guided by Anthropic employees had logged six times the working hours of human operators. Third, employees’ work has shifted structurally, moving away from directly doing tasks and toward building tools, monitoring workflows, and managing agents. Fourth, Anthropic has begun setting up early-warning systems and internal redeployment mechanisms, but the report says company-level responses are still not enough to solve broader social problems such as the narrowing of entry-level opportunities.
As one of the companies at the front edge of AI deployment, Anthropic is presented in the report as a useful sample for observing how AI may change work and hiring in other organizations.
Entry-level hiring narrowed even as the company kept growing
One of the report’s most closely watched data points is the drop in junior technical hiring.
Since early 2025, the share of junior employees in Anthropic’s technical organization has fallen from 21% to 14%. Over the same period, the share of entry-level technical hiring dropped from 27% to 9%. After December 2025, the number of job postings for junior roles declined especially sharply.
This did not come from a contraction in the company. The report says Anthropic tripled in size over the past year, and the absolute number of technical jobs continued to rise. What changed was the mix of hiring: new roles increasingly clustered in mid-level, senior, and specialist ranks, while the entry point for junior talent became much narrower.
According to the report, discussion of lower demand for junior roles first surfaced in core technical teams in January 2026. Some teams that made heavy use of Claude found that routine engineering work, including basic coding and bug investigation, could increasingly be handled by models. The impact was especially visible for performance engineering roles. One interviewee said this almost overnight changed what the team needed from new hires.
By April 2026, engineering and research teams had begun systematically lowering the weight of lower-level roles in hiring plans. In an internal survey that month, about one-third of machine learning managers and one-fifth of software engineering managers said they had reduced hiring because of AI use.
The report links the early pressure on junior roles to the nature of the work itself. Entry-level jobs often involve clearly defined, execution-heavy tasks, which are areas where large language models currently perform relatively well. Coding, data processing, document organization, and customer support were all described as tasks affected to varying degrees by AI, with model capabilities continuing to improve.
In the report’s framing, if model capabilities jump every few months, the bottom rung of the skills ladder may be the first to be displaced.
A similar pattern showed up outside technical teams. In many companies, growth usually brings added headcount in administrative, finance, and HR support functions. At Anthropic, the report says, agent assistants, automated review systems, and outsourced ticket services allowed the company to keep a lean clerical workforce even as it expanded quickly.
One administrative assistant built a personal agent for schedule management and email sorting and said the team had already moved past the stage of using Claude to enhance entry-level administrative roles. The work had shifted toward higher-level coordination and planning.
Agent labor overtook human labor hours
Behind the change in hiring structure is the rapid spread of agents inside the company.
In March 2026, the combined working hours of two categories of AI agents at Anthropic — autonomous agents and agents directly guided by employees — exceeded human employee working hours for the first time, the report says. Those hours kept rising after that. By August 2026, interactive agents directly guided by employees alone had reached six times the working hours of human employees.
Many of those agents were deployed in software engineering, where they handled test-case generation, code review, infrastructure maintenance, and codebase migration. The report notes that this figure does not include fully autonomous agent clusters. For an individual engineer, that means multiple agents can be assigned to run tasks continuously while the engineer reviews their code and analytical output. The engineer’s role, in turn, shifts from writing code directly to assigning tasks and checking results.
Productivity changes also appeared in the numbers. The report says average code commits per engineer increased fivefold over six months. Query throughput for data scientists climbed sharply. Output also rose in non-technical roles, including legal, accounts payable, and sales.
To separate the effect of AI from broader business growth, the research team used a control analysis. It tracked 192 engineers who adopted coding agents and compared them with 74 engineers who had not yet adopted the tool during the same period, estimating the effect with employee and month two-way fixed effects.
The result: code commits rose rapidly after engineers adopted coding agents, and the increase was statistically significant.
The report also examined whether AI generated lower-quality code. Researchers compared, for the same employee, in the same month, and within the same code directory, code produced directly by agents with code written manually by the employee. They used the share of code deleted or reverted within 90 days as the quality measure. The analysis found no statistically significant difference in code quality between the two. At least by the reversion-rate metric, AI-generated code performed on par with handwritten code.
Non-technical roles also saw productivity gains. The report says sales representatives completed more customer transactions after adopting agents. One account executive estimated that Claude automated about 20% of routine work, including drafting emails and operating customer management systems. Another interviewee said the time freed by automation allowed a 30% increase in customer meetings.
From task execution to managing agents
More agents and higher productivity did not remove employees from the workflow. Instead, human attention is becoming the bottleneck.
In interactive sessions, agents spend about 80% of their time waiting for people to issue instructions, confirm outputs, or provide the next step, according to the report.
The research team also converted the tokens generated by agents into reading time using typical human reading speed. It found that the output produced in interactive Claude sessions in non-technical departments was already large enough to consume a substantial share of employees’ weekly working time. Simply reading and processing AI-generated content had become a significant workload of its own.
The report calls the emerging pattern “meta-work” — managerial work organized around agents. That includes building management tools, writing standing instructions, designing monitoring dashboards, and arranging agents to review one another. The core of this work is no longer performing a task directly. It is designing and managing a workflow in which AI performs the task.
Internal data captured that shift. In autumn 2025, meta-work accounted for about 10% of employees’ interactive Claude sessions. By spring 2026, that share had risen to 22%. Most of the increase came from agent-management activity, including writing specifications for multi-agent systems, arranging cross-review among agents, and monitoring workflow status.
The change also showed up in hiring criteria. Through automated coding analysis of interview scorecards covering thousands of Anthropic employees since 2024, the research team found that technical roles mentioning AI tool-use capability in evaluations rose from about 5% to about 15%. For non-technical roles, that share increased from about 7% to about 28%.
At the same time, the weight of different evaluation dimensions shifted. In technical roles, autonomy and judgment were among the fastest-rising criteria, while the relative weight of domain expertise declined. In non-technical roles, customer orientation and mission alignment gained importance, while discussion of execution skills became less prominent. One hiring manager said he expected model capabilities to remain uneven across domains and wanted team members who could move flexibly between them.
One case in the report illustrates that shift. A design team responsible for writing greetings for Claude.ai developed a context graph system in spring 2026 to guide agents in generating product copy, which expanded the team’s output capacity. After that, the team moved into agent prompt engineering and began writing model evaluation content and prompts.
The report summarizes the change this way: writers who once used language to influence human behavior are now using language to influence Claude’s behavior. The job title and outer shell may be unchanged, but the substance of the work has shifted.
The challenge to the bottom rung of the career ladder
The Anthropic case offers a concrete view of how AI can reshape employment structure. As a company at the leading edge of AI application, its internal changes may also foreshadow what other industries face next.
So far, the report says, the first impact has been concentrated in entry-level cognitive work. Junior programmers, entry-level analysts, frontline customer support staff, basic copywriters, and junior finance workers have traditionally been common starting points for university graduates entering the workforce. Those roles are now increasingly affected by AI.
The problem is not just the number of jobs. Entry-level roles are also where workers build experience, learn business processes, and move up over time. If that layer shrinks, opportunities for young people to get a first job and accumulate experience may shrink with it. And if employers increasingly ask newcomers for AI tool fluency, judgment, and autonomy, the question becomes when and how those abilities are supposed to be acquired.
The report discusses several measures companies can take.
- First, use broad-based equity sharing so employees participate in the gains from AI-driven productivity increases. When workers are asked to manage agents and take on more decision-making responsibility, aligning personal incentives with organizational goals becomes more important.
- Second, build early-warning systems to monitor AI’s effect on different roles and plan internal redeployment for affected employees in advance.
- Third, strengthen internal mobility so workers can move into new areas as older tasks are automated instead of being pushed directly into the external labor market.
Even so, the report says internal company responses are not enough to resolve broader social issues. Anthropic, as a frontier AI company, has the resources to offer internal transfers and gain-sharing arrangements. Many traditional firms and their employees may not. Questions such as who helps companies identify role changes and who bears the cost of skill transitions go beyond what any one company can handle.
The report closes with one line: 「A worried and anxious workforce is not a workforce doing its best work.」
That points to more than economics. The report frames AI’s impact on employment as psychological and social as well. When the definition of work shifts quickly, when the bar for entry keeps rising, and when people feel pressure to keep proving they will not be replaced by AI, uncertainty itself becomes a broad social burden.
Anthropic’s experiment is still unfolding. Agent capabilities are still improving, the number of deployments is still rising, and the division of labor between humans and AI is still moving. The report does not claim to know the final shape of that transition. What it does show is that work may not disappear, but the meaning of work, the threshold for entering it, and the order of what is valued are already being rewritten.