The Architecture

  1. PR opens
  2. Multiple agents review concurrently, each focusing on different aspects
  3. Each finding goes to a separate verification agent before surfacing
  4. Verified bugs are ranked by severity and delivered to the developer

The critical design choice: the agent that finds the bug is never the agent that confirms it. This prevents false positives — the single most trust-destroying behavior in automated code review.

The Numbers

Before Code Review: 16% of Anthropic's internal PRs received substantive reviewer comments.
After Code Review: 54%.

The bottleneck was never capability. It was bandwidth. Human reviewers are good; they just can't review every PR at full depth. Multi-agent review scales review bandwidth without scaling headcount.

What This Means for Solo Founders

Solo engineering is missing one thing more than anything else: a second set of eyes. You ship, test, ship again. The review step that catches subtle bugs — the ones that become production incidents three weeks later — gets skipped because there's no one to do it.

Code Review restores that step automatically, on every PR, without requiring another person. For a one-person engineering operation, that's infrastructure that was previously unavailable.