Six agent hours for every human hour: what Anthropic's workforce numbers mean for your junior staff

TL;DR. On October 16, Anthropic disclosed how AI has changed work inside its own company. In August, employees' AI agents worked six times as many hours as the employees themselves. The share of juniors among technical new hires fell from 27% to 9%. For Germany, the IAB labour research institute finds no collapse yet, but a first warning sign, and it shows up in manufacturing of all places. The bottleneck is moving from the work to human attention. And if you stop hiring juniors today, you will be short of experienced people in five years.

What Anthropic disclosed

The report comes from the research team of the Anthropic Institute. It analyses data that hardly any company shows: headcount, the runtime of AI agents, work output, surveys, calendars and the scorecards from job interviews, plus 17 interviews. Anthropic has built its workflows around its own AI further than almost any other business. The authors therefore read the company as an early signal of what could come elsewhere.

In March 2026, the working hours of all agents exceeded those of humans for the first time. In August, the agents that employees steer themselves worked six times as many hours as those employees. Code commits per engineer have increased fivefold in six months. Output rose outside engineering too: lawyers processed more contracts, accounts payable handled more invoices, sales closed more deals.

The flip side is in the same report. Most of the time in a session, the agent is waiting for a person. Reading everything the agents write in these sessions would take more than half the working week, even in non-technical departments. One engineer compares the job to babysitting: every child calls with the same urgency, whether it is about a juice box or a cut finger.

Who gets hired less

Since early 2025, the share of juniors in technical departments has dropped from 21% to 14%. Among new hires, it fell even faster.

From January 2026, the teams that rely most heavily on Claude noticed that writing code and finding bugs could increasingly be handed off. In April, around a third of managers in machine learning said they were hiring less because of AI, in software engineering about a fifth. Administration also grew only slowly, although the company tripled in a year.

New work is emerging instead. The share of sessions in which employees do not do the task itself, but write instructions, have agents check each other and monitor pipelines, rose from 10% to 22%. The number of standing instructions for agents grew from a few hundred to 8,500 within a year. Time in meetings rose by 20%. In the scorecards from job interviews, how candidates use AI comes up three times as often as in 2024 for technical roles, and four times as often for all others. Autonomy and judgement count for more, narrow domain expertise for less. A design team that used to write the greetings in the product now writes tests and instructions for Claude.

The small print

• A special case. Anthropic sells the very agents it measures, and it is in the middle of a boom. The authors write themselves that much of this cannot be transferred to other companies.
• Hours are not output. What is counted is agent runtime, and for humans a flat 40 hours per week per head. The number does not say how much of it is useful.
• More is not the same as better. The engineers who first used fully autonomous agents were already ramping up beforehand. More code does not mean more value. That said, code written by agents was not deleted or reverted more often than code written by hand.
• The remedies do not help juniors. Anthropic uses an early warning system to track which roles are changing, and moves people internally. Anyone who is never hired gains nothing from that. The report says so too.

And in Germany?

Three days earlier, the German Institute for Employment Research published a matching analysis (IAB-Forum, October 13, in German). It is based on data for around 22.3 million full-time employees from 2022 to 2025. There is no general slump in occupations where AI could do many tasks much faster. In fact, employment there was about 1% above 2022 in 2025, while in other occupations it was a good 3% below.

Only among 20 to 29-year-olds does the picture turn. That is not yet significant, and the weak economy weighs more than twice as heavily. In two areas, however, the link is already statistically robust for the young: in manufacturing and in occupations that mostly do not require a degree. That is home ground for many mid-sized companies. In the US, the signal is clearer. According to the Stanford analysis from August, employment of 22 to 25-year-olds in highly exposed occupations is about 19% below the comparison group. There too, mainly because fewer young people are hired, not because of layoffs.

What this means for a mid-sized company

Hardly any mid-sized company works like Anthropic today. The report is still instructive, because it shows where things get stuck once agents really join the work. Not in the computing, but at the person who checks, decides and approves. And in the entry-level tasks that new staff used to learn from. These are exactly the tasks the teams at Anthropic handed to agents first: writing code, finding bugs, logging progress, drafting emails. Whoever cuts this rung saves today and pays later, when the experienced people retire and nobody has grown into their place.

What you should do now

1. Do not let entry-level roles quietly expire. Redesign them: juniors check what agents deliver and learn from real cases along the way, next to experienced colleagues.
2. Measure the new bottleneck. Track how long agent results wait for review and approval. Plan for that time before you introduce more agents.
3. Treat instructions as company knowledge. Standing instructions for agents need an owner, a place to live and a regular review.
4. Write decisions down. Agents only know what is written. At Anthropic, agents kept sending requests for t-shirts and mugs to a person long after their responsibilities had changed, because that name was on an old document.
5. Hire and train differently. Ask in interviews how candidates use AI and check its results. And give your people time to practise steering agents.

Anthropic is an extreme case, but one with open numbers. It shows that AI does not simply take work away, it shifts it: away from doing, towards checking and steering. If you want to see where this shift will reach your business first, we are happy to talk.