On October 16, Anthropic published internal research showing that the share of junior employees among its technical hires has fallen from 27% to 9% since early 2025. Meanwhile, use of AI agents inside the company has surged. In August 2026, the combined runtime of AI directed by employees reportedly reached six times the estimated working hours of the company's human staff. As AI adoption advances, major changes are appearing in both Anthropic's hiring policy and the nature of its employees' work.
That said, Anthropic is not cutting staff. Its headcount has tripled over the past year, and the number of technical hires has also grown. At the same time, AI is increasingly handling code implementation, while humans are expected to give AI instructions, verify its output, and make decisions. As the share of junior hires declines, a new challenge is emerging: how to cultivate the judgment needed to make the most of AI.
The junior hiring share is falling while the company expands
The internal study is a technical report by four authors at The Anthropic Institute, including Nathan Wilmers, analyzing the company's employment and hiring records and AI usage logs. It also examined work output metrics, meeting schedules, and evaluation records from hiring interviews, supplemented by interviews with 17 employees. It is an observational study of a company that develops AI and actively uses it in its own operations, so it is not representative of companies in general. It has also not been published in a peer-reviewed academic journal, which should be kept in mind.
According to the study, juniors fell from 21% to 14% of technical-department employees, and from 27% to 9% of new hires, a sharper drop. Here, "junior" refers to employees at lower internal job levels, not necessarily young people or new graduates. The study does not give specific age brackets, so these figures cannot be read directly as hiring trends for people in their 20s or for new graduates.
Hiring managers' attitudes have also shifted. In January 2026, there was internal debate over whether demand for junior technical hires would shrink as it became easier to delegate writing code and finding bugs to Claude. In an April survey, about one-third of managers in machine learning and about one-fifth of managers in software development said they had reduced hiring because of AI use. It appears that as AI takes over work once done by junior staff, hiring decisions are starting to be affected.
Still, a lower share of junior hires does not mean fewer hires overall. Technical staff as a share of all Anthropic employees also fell, from 61% to 56%, yet the number of technical hires rose. This is because sales, finance, HR, and other departments grew by even larger amounts. In a rapidly growing business, administrative and sales functions also have to expand, so changes in workforce composition cannot all be attributed to AI.
Nor can the published ratios show how much the actual number of junior hires rose or fell. What is clear is that, in Anthropic's fast-growing technical hiring, the proportion of more senior people has risen compared with before.
AI runtime six times that of humans, but not a measure of work done
In March 2026, the runtime of Anthropic's AI agents, autonomous and interactive combined, exceeded estimated human working hours for the first time. By August, the runtime of interactive AI sessions directed by employees, plus the sub-agents launched from them, reportedly reached six times that of humans on its own. This suggests AI has become deeply embedded in internal operations, executing work on a scale well beyond the time humans spend working directly.
However, caution is needed with this comparison. The AI side counts time spent on reasoning, tool use, and similar activity, and excludes waiting periods longer than 30 minutes. The runtimes of multiple sub-agents launched by a parent agent are also added together. Human working time, meanwhile, is not based on actual work records but is an estimate: headcount multiplied by 40 hours a week. Because the runtimes of AIs operating simultaneously add up, AI runtime reaching six times human time does not mean six times as much work was completed.
Anthropic's research examines not only AI usage but also changes in hiring and work output. The main metrics, and what can and cannot be concluded from each number, are summarized below.
| Metric | What was measured | What the number alone does not tell us |
|---|---|---|
| Junior share of technical hires, 27%→9% | Share of juniors among new technical hires since early 2025 | Number of juniors hired, number of layoffs |
| AI runtime 6x that of humans | Combined runtime of interactive AI and sub-agents in August 2026. Human side estimated as headcount × 40 hours/week | Volume or value of work completed, quality comparison with human output |
| Code change records about 5x | Number of change records per engineer compared with six months earlier | Business value of the software, share of the increase due to AI alone |
| Code quality comparison | Compared deletions, reverts, and similar for human- and AI-written code added by the same employee to the same repository in the same month | Overall quality, including long-term maintainability and safety |
These are based on the figures and notes in Anthropic's study, and the measurement periods and units differ by metric. The ability to run AI at scale and the production of valuable output need to be evaluated separately. It is also not possible to calculate a productivity improvement rate from AI runtime, or to work backward to the number of employees who have become unnecessary.
Code change records up about 5x, but not attributable to AI alone
According to Anthropic, code change records per engineer rose about fivefold compared with six months earlier. Salespeople's closed deals and the legal department's contract processing have also increased. A key finding reported in the study is that not only has AI use expanded, but metrics of employee output have changed as well.
However, at a fast-growing company, workload and output can rise regardless of AI. For example, a sudden surge in product demand would naturally tend to increase salespeople's closed deals. To separate out such effects as far as possible, the research team compared 192 engineers who adopted coding agents with 74 who did not during the observation period. Tracking changes for the same employees before and after adoption, and accounting for business conditions in the same month, the analysis found a tendency for the amount of added code to rise after AI adoption.
That said, this analysis is not an experiment that randomly selected who would adopt AI. An engineer planning to begin a large development project may have adopted AI in preparation. In fact, among employees who adopted fully autonomous agents, code output had already been rising sharply before adoption. While AI use and rising output are associated, it cannot be concluded that all of the increase is due to AI.
The team also tested output quality. Comparing code added by the same employee to the same repository in the same month, they found no clear difference between human-written and AI-written code in the number of lines later deleted or the share of changes that were reverted. The analysis used indicators such as code deleted within 90 days and settings that skip tests as measures of quality.
This suggests that even as code output grew, the quality indicators examined showed no notable deterioration. However, code change records can increase simply by splitting work into smaller pieces, so they do not directly represent software value or development efficiency. The study also did not evaluate long-term maintainability, safety, or value to customers. To judge AI's productivity gains, one must consider not only the number of code changes but also how much time humans spend on later verification and correction.
As AI processing capacity grows, human review becomes the bottleneck
As AI use increases, new burdens are also falling on Anthropic's employees. Analysis of interactive AI sessions found that about 80% of elapsed time from start to finish was spent with the AI waiting for human input. This figure, however, is calculated differently from the "AI runtime six times human time" figure above. The 6x comparison excludes long waits, whereas the 80% figure is the share of total session time spent waiting. The two therefore cannot be combined to calculate AI's effective working hours.
The fact that AI is waiting for human input does not necessarily mean employees have less work. The research team also estimated the time humans need to read text generated by AI. Assuming 0.75 words per token and a reading speed of 250 words per minute, even non-technical employees would need more than half of their working week to read all of the AI's output. This is not a measurement of actual reading time, but it illustrates how difficult it has become for humans to review everything in an environment where AI generates huge volumes of output.
To cope with this burden, employees have begun building systems to manage AI's work. Examples include preparing instruction documents for multiple AIs, having another AI review deliverables, and monitoring progress on dashboards. According to the study, the share of sessions concerning such AI management rose from 10% in autumn 2025 to 22% in spring 2026. However, this comes from automated summarization and classification of interactions with AI, and does not mean employees spend 22% of their working hours managing AI.
Alongside the expansion of AI use, time spent coordinating among humans is also growing. Meeting schedule data showed that the median employee's meeting time rose 20% from the previous year, and 16% even for technical staff who are not managers. Company growth and changes in job mix also play a role, so AI cannot be named as the direct cause. Even so, when AI finds a defect, if the background involves problems between departments, it is humans who decide who responds and what takes priority. The faster AI can move work forward, the more likely it is that humans' capacity to review results and make necessary decisions becomes the constraint.
Meanwhile, software creation is spreading beyond specialist engineers. The study describes a case in which a salesperson used Claude to build a dashboard suited to their own work. With AI, employees who know their work well can build the tools they need without asking a specialist department to develop them. But that does not remove the responsibility to verify that the finished tool works correctly and to coordinate with other departments.
Where will the judgment prized in hiring be developed?
The criteria for evaluating candidates in Anthropic's hiring interviews are also changing. In interview evaluation records for technical roles, mentions of autonomy and judgment increased, while mentions of knowledge in specific fields declined in relative terms. Beyond the ability to write code accurately and quickly, abilities such as deciding what to delegate to AI and how to evaluate its output are increasingly valued.
This does not mean specialized knowledge has lost its value, however. A study on Claude Code use that Anthropic published in June analyzed about 400,000 sessions from October 2025 to April 2026. In a typical session, humans made about 70% of planning decisions, such as what to build and what state counts as finished. Claude made about 80% of execution decisions, such as specific implementation steps. This shows that even as AI decides more about how work proceeds, humans still mainly define the goals and completion criteria.
Expertise in that study means not job title or age but how well one understands the problem at hand. For example, even an accountant with no programming experience can use AI as someone with expertise in that task, if they can correctly specify the rules to be reconciled and the exceptions in month-end processing. Conversely, if one cannot spot errors in AI's output, the faster the work proceeds, the greater the risk of overlooking problems. This shift in how AI is used lies behind the growing emphasis on judgment in hiring.
But the more judgment is emphasized, the more the question of how to develop it matters. The study describes a team that had been writing greeting messages and similar text within a product; after introducing a system that has AI generate the wording, the team shifted its work to writing prompts and evaluating models. The company emphasizes identifying changes in job duties early and reassigning employees to new work. Employees who have already built up experience inside the company can move from work AI replaces to work that uses and evaluates AI.
However, the research team itself acknowledges that such measures alone cannot solve the problem of the falling junior hiring share. Internal reassignment supports existing employees and does not benefit people who are never hired. Traditionally, junior engineers have developed the ability to judge design and quality while implementing code and fixing bugs. If AI takes over that work, through what experience will newcomers build judgment and grow into people entrusted with more advanced work? Even if reassigning existing employees succeeds, this question remains.
Outside research also shows changes in junior hiring trends worth watching. In an explanation of the revised version published in August of their study analyzing ADP payroll records, researchers at the Stanford Digital Economy Lab noted that the weakness in employment among young workers in AI-exposed occupations is showing up through fewer hires rather than increased departures. However, a causal link to AI has not been established, and the study does not conclude that large-scale job losses are occurring across the economy. Also, that study uses age as its criterion, whereas Anthropic's research uses internal job level, so the figures cannot be compared directly. Still, both share the point that looking only at the number of layoffs may cause one to miss AI's effects on employment.
Going forward, the focus is on how many juniors Anthropic actually hires, and through what practical experience and mentoring it develops them. It will also be necessary to confirm whether the quality of AI-generated code holds up over the long term. Even if AI makes it possible to process large volumes of work in a short time, failing to develop people who can evaluate the results and make important decisions may eventually lead to other problems. How to raise AI productivity while passing human expertise and judgment on to the next generation looks set to be an important challenge for corporate management.