TLDR Tech / Data
Better tools made Copilot code review worse. Here's how we actually improved it
GitHub added better shared tools to Copilot code review, but its benchmark results got worse: token costs rose and fewer issues were caught because generic tool instructions caused the agent to browse the full repository instead of starting from the diff. Rewriting the instructions to "start from the diff, narrow with grep and glob, read exact evidence with view" cut average review cost by 20% with no quality loss.
Read the full story