GitHub: Better Tools Made Copilot Code Review Worse — Rewriting Prompts Restored Quality at 20% Lower Cost

On July 10, 2026, GitHub published an unusually candid engineering post: when Copilot code review was moved to better-maintained tools, results initially got worse. Analysis showed that the problem was not the tools, but the instructions (agent prompts) that no longer matched the actual workflow of a code reviewer.

Where was the mistake?

The old instructions led the agent into ‘broad exploration’ — scanning a large portion of the repository before focusing on the changes. This spent tokens and time on context that a pull request review generally does not need. When the new, more capable tool received those same instructions, it simply did the wrong thing more efficiently.

How does the diff-first approach work?

The solution was to rewrite the instructions around two principles. First: batch discovery operations (grouped information gathering) before reading individual files, instead of interleaved exploration. Second: analysis anchored to the PR diff — the agent starts from what actually changed, and fetches broader context only when needed. The result is around 20% lower average review cost with unchanged quality.

Why is the lesson universal?

The finding is concrete evidence of the thesis quantified that same week by the arXiv paper ‘Harness Effect’: the economics of AI agents are determined by orchestration (instructions, flow, context), not the model itself. GitHub’s case also illustrates the flip side — upgrading a model or tools without adapting the instructions can make results worse. For anyone building AI coding tools, the practical message is clear: before swapping out the model, read and rewrite the prompts.