- Book: System Design Pocket Guide: Interviews
- The series: Pocket Guides for Developers — 8 short books on agents, RAG, prompts, observability, events, databases and system design — the whole series
- My project: Hermes IDE | GitHub — an IDE for developers who ship with Claude Code and other AI coding tools
- Me: xgabriel.com | GitHub
On October 16, Anthropic published a report on its own workforce, and the sharpest number in it is about people at the start of their careers. The share of junior employees in its technical departments "slipped from 21% to 14% since early 2025," and the junior share of technical hires "fell even faster, from 27% to 9%" (Anthropic, How employment is changing at Anthropic).
This is the company that makes Claude and Claude Code, writing about itself. The report's own opening says why that matters: "As a company at the frontier of AI adoption, Anthropic could be a leading indicator of how AI will shift work organization, hiring, and labor demand across the economy" (Anthropic). Japan's Business+IT reported the same hiring and commit figures on October 17, along with the report's warning that Anthropic may not generalize to the wider economy.
The hiring drop is the headline. For you, the more useful part is what the report says the remaining work looks like, and what interviewers there now write on their scorecards.
The numbers, from the report
All of these come from the report itself:
- Junior share of technical employees: 21% down to 14% since early 2025.
- Junior share of technical hires: 27% down to 9%.
- Technical share of total headcount: 61% down to 56% over the last year, as sales, finance and recruiting grew faster.
- Code commits per engineer: "5x their velocity as 6 months ago."
- Engineers per product manager: roughly 17 to 1, and rising since 2025.
- "Meta-work" such as building agent tools and monitoring pipelines: from 10% of the average employee's interactive Claude sessions in fall 2025 to 22% in spring 2026.
The obvious question about 5x commits is whether the code got worse. The authors checked. They compared code a person wrote directly with code their agent wrote in the same month and directory, and wrote: "we find no difference in rates of lines of code subsequently removed or reverted" (Anthropic). They also say plainly that "More code does not necessarily mean better code," and note that AI errors may be harder to spot than human ones (Anthropic).
The junior share of Anthropic's technical hires fell faster than the junior share of its technical staff. The report describes the trend and does not claim AI caused it.
Agents logged six times the hours. Their sessions mostly waited on you
The report counts agent hours from session telemetry. In March 2026, "the sum of autonomous and interactive agent work hours leapt past human work hours for the first time." Narrowing to agents that individual employees run, "by August the agents worked 6 times as many hours as their human handlers" (Anthropic).
How that is counted matters. Agent hours are runtime, including inference and tool use, with waits longer than 30 minutes left out. Human hours assume a 40-hour week (Anthropic).
The more telling figure points the other way. "Claude sessions amount to around 80% idly waiting for human input" (Anthropic). The report also finds the median employee spent 20% more time in meetings than a year earlier, and in an April 2026 survey, 20% of respondents described a shift from doing the work to managing Claudes (Anthropic).
Put those together and the job described is less typing and more deciding what the agent should do.
What Anthropic's interviewers write down now
The report reads interview scorecards. In 2024, a candidate's track record with AI tooling came up in about 5% of assessments for technical roles. By 2026 that rate had tripled (Anthropic).
Three sentences describe the shift (Anthropic):
- "The clearest change is increased emphasis on whether candidates have demonstrated creative use of AI tools."
- "the single biggest increases were in comments on applicants' autonomy and judgment."
- "Discussion of domain expertise shrank as a share of overall assessment."
My reading, not the report's: when an agent writes most of the implementation, the interview moves to the parts it can't do for you, like deciding what to build, cutting the work into pieces an agent can finish, and catching the diff that passes the tests and is still wrong. If you are interviewing, a concrete story about an agent change you rejected, and why, now says more about you than another solved puzzle.
Read the caveats before you repeat the headline
The authors are careful, and the post you share should be too. From the report:
- The junior evidence is descriptive. "Descriptively, we see junior employment slowing in jobs that use Claude for the most complex tasks." It is not shown to be caused by AI.
- Adoption wasn't random. The productivity estimates compare 192 engineers who adopted coding agents with 74 who did not, and the authors note that "tool adoption is not random." People who moved to fully autonomous agents were already on a steeply upward output trajectory compared with non-adopters.
- Anthropic is unusual, broad stock ownership included, and the authors caution against assuming its practices carry over to the whole economy.
- On commitments, Anthropic says it "is standing up an ongoing early warning system to monitor jobs facing AI-driven transformation" and has "prioritized internal redeployments." Then the report adds, in a line to keep handy when someone asks what the company is doing about it: "none directly address the slowdown in junior hires documented in this report."
When 80% of a session is waiting on you
I build Hermes IDE, so weigh this section with that in mind. The 80% figure describes the problem Hermes is built around: knowing which agent is waiting on you.
In Hermes, every session shows one status: working, needs approval, asked you, done, error or idle. The attention inbox counts the agents blocked on you, and ⌘I (Ctrl+Shift+I on Windows and Linux) jumps to the one that has waited longest. You get a system notification when an agent is blocked or done, never for the session you are already looking at. Each task runs on its own hermes/<name> branch in its own worktree (Hermes README).
Codex on orders-types and Claude Code on orders-api, each in its own hermes/ worktree, with the sidebar showing status, context use, cost and branch.
It doesn't shrink the waiting. You still answer every prompt yourself, because Hermes never answers a permission prompt for you and never types into a terminal on its own (Hermes README). What it changes is that you stop polling tabs to find out who needs you. It is free, source-available under BSL 1.1, and runs on macOS, Windows and Linux (download). It is not affiliated with Anthropic, OpenAI or any agent vendor.
Your turn
This is one company's data about its own staff, so read it as a close-up of one team rather than a labor-market trend.
If you hire engineers, has your scorecard changed this year, and does it now ask about judgment with an agent in the loop? If you are early in your career, what is the strongest proof of judgment you can show an interviewer? Tell me in the comments.
If this was useful
If interviewers are grading judgment, the place you show it is a design round: requirements, trade-offs, and why you picked one option over another. The System Design Pocket Guide: Interviews walks through 15 real system designs step by step, with that reasoning written out. It is part of Pocket Guides for Developers, 8 short books on agents, RAG, prompts, observability, events, databases and system design. And if you run more than one coding agent, Hermes IDE shows you which one is waiting on you.
