Anthropic: AI Agents Now Outnumber Employees 6 to 1, and Junior Hiring Is Shrinking

Anthropic has published two reports on how AI is changing work. One describes the changes inside the company itself; the other compares the expectations of its employees with those of ordinary Americans. The headline numbers: by August 2026, Anthropic has six times more AI agents than people, and back in March the agents for the first time worked more hours than the employees. Productivity has risen as well. In parallel, the company compared its employees' forecasts with a Morning Consult survey of nearly 11,000 American adults. The gap in expectations turned out to be huge, even though the two groups hold similar economic models.

How hiring has changed inside Anthropic

Over the past year, the company has noticeably shifted the structure of its hiring. The share of junior openings has fallen, as has the share of employees in technical roles. This is happening at a lab that builds automation tools itself and uses them earlier than most.

Anthropic measures the effect on output in several ways. Code volume per person has grown, but the company itself calls this metric notoriously imperfect, meaning flawed by design. So the report cites other indicators too: each sales manager has closed more deals and processed more contracts. The researchers also tracked the same employees over time and found that a person's output rises with how deeply they use AI tools.

The flip side of the speedup is that human attention becomes the bottleneck. Employees spend more and more time building systems that help them supervise the agents' work. In effect, part of the labor is shifting from carrying out tasks to overseeing those who carry them out.

The expectations gap

The second report compares a survey of Anthropic employees with an August survey of Americans, weighted to reflect the adult population. Both groups expect rapid progress, but Anthropic employees expect it to come even faster.

Key numbers:

• 39% of Americans and 91% of Anthropic employees expect AI to be able to make a Nobel-level discovery by 2030.
• More than half of Anthropic employees believe that by 2030 AI will be able to do at least 80% of their current work.
• Among Americans, only 5% expect that level of capability.
• Two in ten Anthropic employees hold views close to those of the typical American.
• Within each group, the variation is large.

Different experiences, one model

The most interesting finding concerns the reasons for the divergence. The authors argue that employees and the public do not hold different theories about AI's effect on employment. What differs is experience: people encounter the models' real capabilities in different ways. And the same relationship holds in both groups: those who expect AI to do more of their own work also predict a larger rise in unemployment.

This is an important caveat for any debate between "optimists" and "pessimists." If the finding holds, expectations are more likely to converge as more people gain direct experience with frontier systems than as new arguments appear.

The limitations deserve separate mention. Anthropic openly acknowledges that an AI lab differs from an ordinary business in how it adopts AI. Moreover, this is a self-report from an interested company, and its employees are by definition selected for their closeness to the technology. Still, as an early indicator the data is useful: it shows where other organizations may end up, including future AI-native companies.

What companies should do

Based on Anthropic's experience, executives should rethink three things. First, the hiring pyramid: junior tasks are the first to go to agents, so newcomers need a clear growth path. Second, metrics: code output reflects results poorly, and it is more reliable to look at closed deals, processed contracts, and the trajectory of the same people over time. Third, investment in monitoring: when agents outnumber people many times over, human attention needs tools for observation and verification, or it will be what limits growth.