Anthropic Clocks Its Own Engineers: The Results Are Weird

Anthropic did something that takes either confidence or nerve: they studied themselves. Not in a PR-flattering way — actual surveys, 53 in-depth interviews, internal usage logs, the whole epidemiological apparatus turned inward on 132 of their own engineers and researchers. What they found is a genuinely strange portrait of what it looks like when a workplace runs hot on AI assistance before most workplaces have even warmed up.

The full writeup is worth reading, but here's the shape of it.

The Numbers That Sound Good

Engineers self-report using Claude in 60% of their work and claim a 50% productivity boost — roughly 2-3x what they reported a year ago. A quarter of that Claude-assisted work (27%) consists of tasks that wouldn't have happened otherwise: dashboards nobody would've commissioned, exploratory work too expensive to justify manually, the nice-to-haves that quietly rot on every team's backlog. That's not just going faster. That's a different surface area of work getting done.

On the Claude Code side, the agentic run length doubled in six months — from ~10 autonomous actions before needing human input to ~20. Usage shifted toward genuinely complex tasks: implementing new features went from 14% to 37% of Claude Code use. The tool is doing more, and doing harder things.

The Numbers That Sound Complicated

Here's where it gets interesting. Engineers can "fully delegate" 0-20% of their work to Claude, despite using it constantly. The rest involves active supervision and validation — Claude as collaborator you can't yet trust to go handle things unsupervised. One engineer described the delegation logic plainly: "The more excited I am to do the task, the more likely I am to not use Claude." So the AI gets the boring stuff, and the human keeps the craft. For now.

Skillsets are broadening — security engineers reading unfamiliar code, alignment researchers building front-end visualizations, backend folks not flinching at the database layer. Everyone's going full-stack via vibes and Claude. But the corollary is that some depth is thinning. As one engineer put it: "When producing output is so easy and fast, it gets harder and harder to actually take the time to learn something." You can have breadth or you can have the struggle that produces depth. Right now you're getting breadth, fast.

The Social Texture Nobody's Talking About

This is the part that stuck with me. Claude is now the first stop for questions that used to go to colleagues. Senior engineers are getting fewer questions from juniors. Mentorship is declining not because anyone decided it should but because the friction that created it — the necessary ignorance that made you walk down the hall — is getting routed around. One person said it plainly: "I like working with people and it's sad that I 'need' them less now."

That's not a productivity metric. That's a texture-of-work metric, and it's not improving.

What This Is Actually About

Anthropic is right to flag their own position here — these are privileged workers at a stable company with early tool access, and the findings don't transfer cleanly to a warehouse or a law firm. But they're also right that this is likely a harbinger. The specific shape of what's happening — breadth up, depth uncertain, collaboration patterns disrupting quietly, career trajectories genuinely unclear — is probably going to look familiar in other knowledge-work contexts, just on a delay.

The engineers who said only that it was "hard to say" what their roles might look like in a few years? That's the honest answer. The ones expressing short-term optimism and long-term existential dread in the same sentence are probably reading the situation accurately. Both things can be true. Usually are.