GitHub swapped Copilot code review's custom exploration tools for the shared Copilot CLI tools, grep, glob, and view, expecting a clean infrastructure win. Instead, benchmarks showed higher review costs and fewer issues caught. The tools were not broken. The instructions were.
Traces revealed the agent was browsing repositories like it was doing general exploration, widening searches, guessing paths, accumulating context, instead of starting from the diff and asking targeted questions. Every tool result lands in the context window and stays there. The old custom tools automatically bundled surrounding code lines into results, which matched how earlier models behaved when making fewer tool calls. The new tools did not do that, and the instructions never compensated for the difference.
The fix was rewriting the workflow instructions to match how a pull request reviewer actually reads code: start from the diff, pull the minimum context needed to answer a specific question, stop when you have enough. That change flipped the regression into a 20% reduction in average review cost with no drop in review quality. The full writeup is worth reading for the benchmark methodology and the before and after trace comparisons, which make the browsing loop failure mode concrete.
[READ ORIGINAL →]