Design a Code Review Agent

A code review agent is the evidence-grounded inspection question: find actionable defects in a proposed code change, explain them precisely, and avoid wasting the author's attention. Meta asks it in agentic system design, with emphasis on which work needs an LLM, how generated findings are validated, how orchestration and memory work, which tools are available, and what inference costs.

We follow the Interview Framework with evaluation pulled into Phase 1. Announce that choice: a review agent needs an agreed definition of a useful finding before its architecture can be assessed. The Ticket-to-PR Coding Agent owns the shared repository retrieval and sandbox foundations; this walkthrough develops reviewing an existing diff.

All scale, budgets, and operating policies below are design assumptions for the worked answer.