Agentic Code Review
Focuswhat AI review catches well and what it misses; review architectures and custom rules; fitting AI review into a team's pull-request workflow.
You buildA severity-ordered review rubric wired to pull requests, measured on five PRs
Core material≈ 3 h 30 min
- 1
- 2
- 3Talk10 minAI-powered entomology: lessons from millions of AI code reviewsTomas Reimers of Graphite
, with Graphite's written AI code review implementation and best practices (20 min).
- 4Review architecture16 minHow to Kill the Code ReviewAnkit Jain
A five-layer trust model combining specifications, reusable guardrails, deterministic checks, executable test plans, previews, and human alignment.
- 5What automation cannot replace20 minUnderstanding Is the New BottleneckGeoffrey Litt
Review as architectural understanding, mentorship, and coordination rather than only correctness checking.
- 6Integrations15 minAbout GitHub Copilot code reviewGitHub
And Anthropic, Claude Code GitHub Actions (15 min) — the two most common ways to put a reviewer agent on a pull request.
From Cognition
- Video32 minDeepWiki: The GitHub EncyclopediaLatent Space with Cognition
Codebase understanding as a product; the same understanding a reviewer needs.
Deeper material
- LinkAI-Assisted Assessment of Coding Practices in Modern Code ReviewResearch paper
And Google Research, Resolving code review comments with ML.
- Rules and feedback loops10 minYour Coding Agent Doesn't Always Follow Your Rules
Hooks, deterministic checks, asynchronous verification, reviewer agents, and LLM-as-judge tradeoffs.
Build
Create a review rubric ordered by severity: correctness, security, data loss, concurrency, compatibility, tests, maintainability, then style. Require the reviewer agent to cite exact lines, explain the failure scenario, and propose the smallest fix. Use a model/session that did not write the change. Wire it to your pull requests with one of the integrations above.
Evaluate the reviewer on at least five pull requests, including one deliberately seeded with subtle bugs. Label every comment true positive, false positive, duplicate, or low-value. Track acceptance rate and escaped defects. A human still owns merge approval.