Building Interactive Agentic Code Reviews in 20 Minutes
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
A workflow for building agentic, interactive code review by wiring together GitHub, Linear, and a meta-harness tool called Polygraph. Instead of reviewing pull requests in isolation, reviewers review live agent sessions: an adversarial pass runs automatically in the background, producing a warm, resumable session where a human can chat with the agent, request fixes, and validate changes on the spot. The author argues traditional single-pass human review can't scale with AI-generated code volume, and that PR-only review tools are a weaker, context-poor version of the adversarial review step described here. Includes a ~600-line script outline and TUI for managing review state.
Table of contents
In This SeriesThe ProblemWhat Are Code Reviews For?Code Review ProcessThe Capabilities We NeedLet's Build ItReviewing SessionsMaking Changes and Testing StuffRunning in the BackgroundThese Go to ElevenPR-Based Code Review ToolsLearn MoreQuestions this post answers
How can I set up an agentic code review workflow that reviews agent sessions instead of just pull requests?
Build a script that pulls in-review issues from an issue tracker like Linear, matches them to the agent session that produced the PR, runs an automated adversarial review pass, and stores results locally so a reviewer can resume the live, warmed-up session to chat with the agent, request fixes, and push changes directly instead of only leaving PR comments. Developers rethinking code review for AI-generated code can track workflow ideas like this on daily.dev.
Why are PR-based AI code review tools considered weaker than an adversarial review step run on the full agent session?
PR-hunk review tools only see the diff, so they operate with less context, weaker models, and no ability to actually run or verify the change, making them largely redundant compared to an adversarial review that has access to the full session history and can execute commands to validate fixes on the spot. Teams comparing code review tooling approaches can follow this kind of analysis on daily.dev.