Itamar Friedman separates code review into two responsibilities: validating safe, maintainable changes and helping developers align and learn. He argues that automating pull-request review requires preserving both functions, while recognizing that teams differ in their tolerance for production bugs and their willingness to prioritize delivery speed.
Itamar Friedman identifies missing organizational context as a central obstacle to reliable AI review. Scattered instruction files, team documents, conversations and developer experience can produce inconsistent judgments across coding tools. He proposes codifying that knowledge in a shared context system with interfaces that help humans audit decisions and agents use the same standards.
Itamar Friedman extends this context beyond written rules to software architecture, service contracts, incident history and the discussions behind past fixes. A graph of repositories and their dependencies could help reviewers identify when a change breaks another service or conflicts with another pull request. He argues that rules for approving or blocking changes should also be explicit and inspectable.
Itamar Friedman recommends introducing automation gradually, tracking which rules are used, whether their findings are useful and when standards need updating. His proposed direction shifts human attention from inspecting every change toward governing the knowledge and relationships used by review systems. The talk presents a design argument rather than independent evidence that human review can already be removed safely.
Watch on YouTube
