GitHub gave Copilot’s code review agent better tools — and review quality dropped. The fix wasn’t rolling back. It was restructuring the agent’s workflow around pull request evidence using shared Unix-style code exploration tools.
The GitHub engineering team documented the counterintuitive result:
- Migration to shared Unix-style tools — Copilot code review moved to code exploration tools built around Unix conventions rather than bespoke review-specific utilities
- Workflow reshaped around PR evidence — the agent now grounds its review in what the pull request actually changes, not what it assumes the codebase looks like
- Review cost reduced — the new approach cut the cost of running automated reviews
The paradox: upgrading the agent’s tooling degraded its output. More capable tools gave the model more paths through the codebase, which pulled focus away from the actual diff. The fix was constraining the workflow — anchoring the agent to the pull request as the primary evidence source and using Unix-style tools as a shared, predictable interface rather than a sprawling toolkit.
NOTE
This aligns with GitHub’s broader Copilot architecture work. A separate post describes cutting VS Code Copilot Chat from hundreds of MCP tools to a 13-tool core using embedding-guided tool routing and adaptive clustering. Same thesis: fewer, sharper tools beat more, broader ones.
The takeaway for anyone building agent workflows: tool quality matters less than tool-to-task alignment. GitHub’s Copilot code review got better and cheaper when they stopped adding capabilities and started shaping the workflow around the evidence that actually matters — the pull request.