What Do Coding Agents Actually Change? A Taxonomy of 1,254 PR Diffs

Source: arXiv — 2026-07-06

Summary

Illia Dovhoshliubnyi and coauthors classify 1,254 diff hunks from 216 real performance-improving pull requests across five coding-agent systems using an 18-category mutation taxonomy. They find agents overwhelmingly favor name modification, object creation, and type changes — a profile markedly different from the mutation patterns seen in prior human-written or genetic-improvement code corpora.

Key Takeaways

• 1,254 diff hunks, 216 real PRs, five different coding-agent systems, classified into an 18-category mutation taxonomy.
• Agents disproportionately favor name modification, object creation, and type changes over more structurally invasive edits.
• The resulting profile differs markedly from mutation patterns in human-written code and prior genetic-improvement research.
• Useful empirical grounding for "what do coding agents actually do to a codebase" — relevant to both SE practice and search-based software engineering research.

Discussion

(No questions yet — ask follow-ups via a Claude Code chat session on this repo; answers get appended here.)