The skill to hire for in 2027: reviewing

At Anthropic, agent sessions spend 80% of their time waiting for a human. The bottleneck is no longer writing the code. It's reviewing it.

The report Anthropic published on its own teams is pretty clear. Agents there log six times more working hours than employees. "Meta-work" (writing instructions, having one agent check another, monitoring all of it) went from 10% to 22% of sessions. And reading everything their agents produce would take more than half the week.

In their interview rubrics, judgment and autonomy weigh more. Narrow expertise, less.

The rare skill is therefore review. Reading fast, reading right, spotting what's off in a perfectly plausible diff, and knowing how to say no.

Almost no interview process tests it. We still have someone write an algorithm on a whiteboard who will spend their days reviewing an agent's.

And almost no sourcing looks at it. We count commits, we list stacks. Yet on GitHub, a developer's public reviews often say more than their commits: what they notice, what they let through, how they argue a rejection in someone else's repository.

When I need to size up an engineer, that's where I look first.

If your technical interview doesn't include a single code review, you're assessing a job that is disappearing.