Anthropic releases internal employment report: AI agents work 6 times longer than humans; the "meta-work era" has arrived.

This articleis written by Tencent Technology.

On October 16, local time in the United States, Anthropic released an internal research report on its own workforce structure, titled "How Employment is Changing at Anthropic".

This study, based on Anthropic's internal employment records, AI usage telemetry data, employee output metrics, employee surveys, internal communication records, calendar data, and hiring decision scorecards, combined with 17 in-depth employee interviews, systematically analyzed the changes in job content, hiring structure, and workforce needs after the company's deep application of AI.

The report presents four main findings:

First, Anthropic's hiring of entry-level technical positions has decreased significantly, from 27% at the beginning of 2025 to 9% in 2026. However, the company's overall employment scale is still expanding, and it is the entry-level positions that are being squeezed, rather than the total number of positions.

Secondly, the working hours of AI agents are rapidly increasing. As of August 2026, the working hours of interactive agents directly guided by Anthropic employees have reached six times that of human operators. Productivity has improved in multiple roles, and the research team further evaluated the independent contribution of AI through comparative analysis.

Third, the content of employees' work is undergoing structural changes. The focus of employees' work is gradually shifting from directly executing tasks to building tools, monitoring workflows, and managing intelligent agents. At the same time, human attention is becoming a new bottleneck, and recruitment standards are beginning to shift from domain expertise to autonomy and judgment.

Fourth, Anthropic has begun establishing an early warning system and internal staffing mechanisms. However, the report points out that corporate-level responses are still insufficient to address broader social issues such as shrinking entry-level positions.

As one of the companies at the forefront of AI application, these changes within Anthropic provide a model for observing how AI may affect the way other companies work and their hiring needs.

Recruitment of entry-level positions is being reduced.

The most striking data in the report is the significant drop in the proportion of junior technical positions being recruited.

Since the beginning of 2025, the percentage of junior employees in Anthropic's technology department has decreased from 21% to 14%, while the percentage of junior technical job postings has decreased from 27% to 9%. The number of junior job postings saw a particularly significant decline after December 2025.

Anthropic sees a decline in hiring for entry-level technical positions.

This change did not stem from company downsizing. On the contrary, Anthropic has tripled in size over the past year, and the absolute number of technical positions continues to grow. What has truly changed is the hiring structure: new positions are more concentrated at the mid-to-senior level and expert level, while entry-level positions have significantly narrowed.

Discussions about entry-level job requirements first appeared in the core technology department in January 2026.

Some teams that heavily use Claude have found that routine engineering tasks such as basic code writing and vulnerability checking can increasingly be handled by the model. The impact is particularly pronounced on performance engineers. One interviewee stated that this almost overnight changed the team's talent needs.

Claude employs more complex roles, and junior staff experience relatively slow growth.

By April 2026, engineering and research teams had begun systematically reducing the hiring weight of job titles. An internal survey that month revealed that approximately one-third of machine learning managers and one-fifth of software engineering managers reported reducing their hiring due to the use of AI.

Entry-level positions are the first to be impacted, which is related to the characteristics of these jobs.

Entry-level tasks typically feature clear rules and strong execution requirements, which are precisely the areas where current large language models can handle well. Tasks such as coding, data processing, documentation, and customer support are all being influenced by AI to varying degrees, and the models' ability to perform these tasks is continuously improving.

As a model's capabilities advance to a new level every few months, the first thing to be replaced may be the "bottom" of the skill ladder.

Non-technical departments exhibit a similar trend. Typically, as companies expand, support departments such as administration, finance, and human resources also increase their staff. However, at Anthropic, intelligent assistants, automated approvals, and outsourced ticketing services have enabled the company to maintain a lean clerical workforce while rapidly expanding.

An administrative assistant developed a personal assistant agent for schedule management and email categorization, and stated that the team has moved beyond leveraging Claude to enhance entry-level administrative roles, and the focus is now shifting to higher-level coordination and planning tasks.

Intelligent agents achieve 6 times the working hours of employees

Behind the changes in recruitment structure is the rapid popularization of intelligent agents within companies.

In March 2026, the cumulative working hours of two types of AI agents within Anthropic—one autonomously operating and the other directly guided by employees—exceeded those of human employees for the first time. Since then, the working hours of these agents have continued to grow. By August of this year, the working hours of the employee-guided interactive agents alone had reached six times that of human employees.

Intelligent agents have been working longer than humans.

Numerous intelligent agents are deployed in the field of software engineering, undertaking tasks such as test case generation, code review, infrastructure maintenance, and code repository migration. This data does not yet include fully autonomous agent clusters. For an individual engineer, this means they can simultaneously schedule multiple agents to continuously execute tasks and review the generated code and analysis results. The engineer's work is also gradually shifting from writing code themselves to task allocation and result review.

Changes in productivity are also reflected in specific data. The report shows that the average number of code submissions per engineer increased fivefold within six months, data scientists saw a significant increase in query processing volume, and output from non-technical positions such as lawyers, accounts payable, and sales also increased.

After adopting AI technology across multiple functional departments at Anthropic, employee productivity surged.

To eliminate the interference of factors such as business growth, the research team adopted a control analysis method. They tracked 192 engineers who used coded agents and used 74 engineers who had not yet adopted the tool during the same period as a control group, using two-way fixed effects of individuals and months for estimation.

The results showed that after adopting coding agents, the number of code submissions by engineers increased significantly and rapidly.

Anthropic engineers saw increased productivity after adopting coding agents, and the regression rate of AI-generated code was not significantly different from that of handwritten code.

The report also analyzed the question of whether AI would generate low-quality code.

The research team compared code generated directly by an AI agent and code manually written by the employee within the same month and code directory, using the percentage of code deleted or rolled back within 90 days as a quality assessment metric. The results showed no statistically significant difference in code quality between the two methods. At least in terms of code rollback rate, AI-generated code performed comparably to human-written code.

Productivity in non-technical roles also improved. The report shows that sales representatives completed more customer transactions after adopting the intelligent agent. One account manager estimated that Claude automated about 20% of routine tasks, including drafting emails and operating the customer management system.

Another respondent said that the time it took for automation to be released allowed him to increase his number of customer meetings by 30%.

From executor to agent manager

The increase in the number of intelligent agents and the improvement in productivity have not freed employees from their work. On the contrary, human attention is gradually becoming a bottleneck for the output of intelligent agents.

The report shows that in interactive sessions, the agent spends about 80% of its time waiting, requiring humans to issue instructions, confirm results, or provide directions for the next step.

The research team also converted the tokens output by the AI agent into reading time, based on the average person's reading speed. The results showed that the content generated by interactive Claude conversations in non-technical departments was enough to occupy a significant portion of employees' weekly work time. Simply reading and processing AI-generated content has become a considerable workload.

Reading Claude's output alone is equivalent to a full-time job.

Faced with this change, employees have gradually developed a new way of working. The report calls this "meta-work," which refers to managerial work centered around agents. This includes building management tools, writing standing instructions, designing monitoring dashboards, and arranging for different agents to review each other. The core of this type of work is no longer directly executing specific tasks, but rather designing and managing processes for AI to complete those tasks.

Related data also documents this shift. In the fall of 2025, "metawork" accounted for approximately 10% of employee-interactive Claude sessions. By the spring of 2026, this proportion had risen to 22%. The growth was primarily driven by agent management activities, including writing specifications for multi-agent systems, arranging cross-agent reviews, and monitoring the operational status of workflows.

Anthropic employees are increasingly managing agents rather than doing the actual work themselves.

The changes in work methods are also reflected in recruitment standards.

The research team conducted an automated coding analysis of interview scorecards from thousands of Anthropic employees since 2024 and found that the proportion of assessments mentioning AI tool usage ability in technical positions increased from about 5% to about 15%, while in non-technical positions it increased from about 7% to about 28%.

At the same time, the focus of recruitment assessments has shifted. In technical roles, autonomy and judgment have become rapidly growing assessment dimensions, while the weight of domain expertise has relatively decreased. Non-technical roles place greater emphasis on customer awareness and mission orientation, with less discussion of execution skills. One hiring manager stated that he anticipates model capabilities will continue to vary across different areas, and therefore hopes team members can flexibly switch between different domains.

One case in the report also illustrates this change.

In the spring of 2026, an Anthropic design team responsible for writing greetings for Claude.ai products developed a context graph system to guide agents in generating product copy, thereby expanding the team's output capabilities. Subsequently, the team shifted its focus to agent prompting engineering, taking charge of writing model evaluation content and prompts.

The report summarizes this as follows: Writers who once influenced human behavior through language are now beginning to influence Claude's behavior through language. In other words, the title and outward appearance of the position have not changed, but the core of the work may have been different.

The entry-level ladder faces challenges

Anthropic's case illustrates the impact of AI on employment structure. As one of the companies at the forefront of AI application, the changes occurring within the company provide a reference for observing potential future changes in other industries.

Currently, AI's impact on employment is primarily seen at the entry level of cognitive labor. Positions such as junior programmers, entry-level analysts, frontline customer service representatives, basic copywriters, and junior accountants, which were typically the starting point for university graduates entering the workforce, are now increasingly affected by AI.

This change brings more than just a shift in the number of positions. Entry-level positions are also crucial starting points for employees to gain experience, learn business processes, and gradually advance their careers. As this level shrinks, young people may have fewer opportunities to get their first job and accumulate professional experience. If companies increasingly require new employees to demonstrate proficiency in AI tools, sound judgment, and autonomy, then the question of at what stage and through what means these abilities should be acquired becomes a pressing issue.

The report discusses several measures that businesses can take:

First, through a broad equity-sharing mechanism, employees can share in the benefits of AI-driven productivity improvements. As employees need to manage intelligent agents and assume greater decision-making responsibility, aligning their personal interests with organizational goals becomes even more crucial.

Second, establish an early warning system to continuously monitor the impact of AI on different positions and plan for the internal reassignment of affected employees in advance.

Third, strengthen internal mobility mechanisms. As existing tasks are gradually automated, help employees transition to new work areas, rather than directly pushing them into the external labor market.

However, internal measures within companies are insufficient to address broader societal issues. Anthropic, as a cutting-edge AI company, has the resources to offer internal job transfers and profit-sharing mechanisms for its employees. For most traditional industries and their employees, these resources and institutional arrangements may not be available. Who will help these companies identify job changes? Who will bear the costs of employee skills transition? These questions exceed the capacity of any single company to address.

One sentence at the end of the report is worth noting: "An anxious and worried workforce is not a hardworking workforce."

This statement suggests that the impact of AI on employment is not only economic, but also psychological and social.

When the definition of work is changing rapidly, when the barriers to entry are constantly rising, and when everyone needs to constantly prove that they will not be replaced by AI, this continuous uncertainty itself becomes a widespread social and psychological burden.

Anthropic's experiments continue. The capabilities of intelligent agents are constantly increasing, the number of applications is growing, and the division of labor between humans and AI is constantly evolving. The final form of this transformation remains unclear, but one thing is certain: jobs may not disappear, but the meaning, barriers to entry, and value ranking of jobs are being rewritten.

Source:Global Cybersecurity Alliance (GCSA)
Website:www.gcsa.org