How Work Is Changing Inside Anthropic (#AI)

I read the article How Employment Is Changing at Anthropic, published on October 16. Much of it connects with our 3 AImigo episode on the AI-native company and my longread: agents speed up execution, and then processes, roles, and decision-making rules have to change. What interests me here is how that looks inside Anthropic itself.

The authors combined internal data on hiring, AI use, work output, meetings, and communication with 17 interviews. The study has four connected threads.

🔸 Who gets hired and where people move. The junior share of new technical hires fell from 27% to 9%. That is a share of hiring, not a layoff count. Non-technical employees build their own tools, managers commit code more often, and entire teams change their work: after automating product copy, its writers moved into prompts and evals.

🔸 What happens to productivity. Commits, contracts processed, and deals closed per employee are increasing. The authors compare employees’ work before and after adopting agents with that of people who have not yet adopted them. It is an interesting design, but the groups did not form randomly: a major new project could both increase output and prompt someone to adopt agents.

🔸 What the working day becomes. Human attention becomes a bottleneck. The share of interactive Claude sessions devoted to meta-work—instructions, checks, and agent monitoring—rose from 10% to 22%. That is a share of sessions, not working time. Even technical individual contributors spent 16% more time in meetings than a year earlier. Agents need coordination, and people continue to work things out with one another.

🔸 What employers can do. Share the gains from automation, identify changing roles early, and move employees into new work. The authors acknowledge that protecting existing employees does not help people who never get hired. The labor market will need solutions beyond a single company.

Now for my notes in the margins of the attached PDF.

1️⃣ Output still has to become an outcome—a point I explored in “Token Burn as Proof of Work”. More code alone does not tell us what improved for the customer. Even rising revenue per employee does not isolate the effect of AI from business growth. This is a study of one rapidly growing company that sells the very technology being studied.

2️⃣ Meta-work fits our discussion of Loop Engineering in Research Insights. You can spend all day servicing a queue of questions from agents. Or you can design the loop: task, execution, verification, correction, and escalation to a person. I want to know whether that reduces the human attention needed per accepted result. An increase in the number of instructions alone tells us little.

3️⃣ The convergence of roles is also interesting: managers become more hands-on, while engineers get teams of agents. That helps explain the demand for autonomy and judgment. We discussed how to assess them in “AI Hires AI” (post with episode materials). But where will beginners acquire that judgment if the tasks that used to build experience have already been automated? That is the question behind our episode “The Junior Engineer Without Easy Tasks”. And how do you manage all this when the real decisions stay in private messages and agents get outdated documents?

4️⃣ The part that matters most to me is incentives. Shares in a rapidly growing Anthropic give employees a stake in its success. How does that bargain transfer to an ordinary company? Internal mobility can also exist on paper while becoming a stop on the way to dismissal. People need real new roles and a viable path into them.

If I automate my own work, what share of the gains will I receive, and what will I do next? I would ask that before setting employees yet another AI adoption KPI.

#AI #Research #Agents #Engineering #Management