Code review with agents

Two practitioner patterns for agent-assisted review that beat the generic single-prompt reviewer.

• Self-improving reviewer. Run the reviewer on merged PRs, cross-reference developer responses (pushed back, fixed, ignored), classify each comment, find recurring dud patterns, and have it edit its own review prompt on a schedule. The prompt evolves on data, not memory; every change is traceable.
• Multi-lens specialists. An orchestrator analyzes the diff and routes to only the specialists needed (frontend, backend, DevOps), running them in parallel and synthesizing. A Terraform-only PR skips the frontend agent, so you pay only for the expertise you use.

Why multi-agent beats single: token usage explains most of the performance variance, so spending tokens across focused contexts outperforms one generalist with a wall-of-text prompt. See OpenCode for the harness the multi-lens writeup is built on, and Claude Code for the self-improving reviewer.

Sources: 0010-self-improving-code-review.md, 0011-multi-agent-review-opencode.md

Edit wiki/code-review-with-agents.md; see how we maintain the wiki.