Better tools made Copilot code review worse. Here's how we actually improved it.

What this post added

This post details the engineering effort to improve Copilot code review by optimizing the instructions for shared code exploration tools (grep, glob, view). It describes how the initial migration to these tools led to a regression in performance and effectiveness, characterized by a 'browsing loop' rather than focused investigation. The core contribution is the redesign of tool instructions to align with a reviewer's workflow: starting from the diff, narrowing the search with grep and glob, and using view only for specific evidence. This involved batching discovery and reads, and refining error recovery strategies. The post emphasizes the use of internal benchmarks to debug agent behavior and validate the effectiveness of these instruction changes, resulting in approximately 20% lower average review costs.