Copilot code review: Unix tools over bespoke

Specialized AI tooling for code review introduced more friction than it removed.

What changed

• Copilot's code review workflow migrated from a bespoke, AI-centric tool to shared Unix-style code exploration utilities.
• Agent workflows were refocused around evidence found within pull requests, rather than solely on AI-generated suggestions.
• Review cost was reduced by approximately 30% through this shift.

Why it matters

This post details a pragmatic retreat from over-engineered AI tooling. The initial approach, aiming to streamline code review with dedicated AI features, paradoxically increased overhead. By migrating to simpler, composable tools akin to traditional Unix utilities, GitHub engineers found they could more effectively integrate AI assistance into existing developer workflows. The honest version: what seemed like a feature-rich solution was actually a productivity bottleneck. This shift suggests that for complex tasks like code review, composability and integration with established developer habits may outweigh bespoke, AI-first solutions. It's a reminder that sometimes the best AI tool is one that doesn't feel like an AI tool at all.

The catch

While the move to shared tools reportedly reduced review cost by ~30%, the announcement does not specify the baseline cost or the exact metrics used for this calculation. Furthermore, the success hinges on the quality and availability of the underlying