Parsimony asks a useful question alongside whether a coding agent solved a task: how much implementation code did its fix add? It analyzes published benchmark patches without running the target code, counting syntax elements such as statements, calls, names, and literals rather than lines. Deletions subtract from the net total; squeezing code onto one line does not earn a better result.

  • For developers evaluating agents: The browser boards show code footprint alongside solve rate and measurement coverage. Current boards compare each model with others on the same solved tasks, reducing the advantage of solving only small tasks.
  • A verified recent release: The October 11 Go beta release corrects its parser and remeasures the full DeepSWE Go cohort. Six previously unmeasurable attempts become measurable; previously measured primary footprint counts stay unchanged. This is a meaningful analyzer/data update, not a brand-new project.
  • Use it: Browse the boards without installation, or clone the repository and use its Python CLI. Python analysis requires Python 3.12+ with no runtime dependencies; JavaScript/TypeScript require Node.js, and Go needs additional parser dependencies and a C compiler.
  • Bring your own results: The custom benchmark viewer accepts a JSON file containing model, task, amount, and optional solved status. It stays in the browser tab. These supplied counts are unverified and do not become official leaderboard entries.
  • Cost and access: The analyzer is MIT-licensed open source; the public boards are directly accessible. Measuring existing patches does not require new paid model calls. Generating your own agent runs can still incur model and infrastructure costs.

The limitation: Smaller is not necessarily clearer, safer, or more correct. Parsimony trusts upstream pass/fail results rather than rerunning tests, excludes tests/docs/generated code from its implementation counts, and cannot certify maintainability. Keep languages and analyzer versions separate when comparing results. Treat it as another review signal—not a reason to prefer a tiny brittle fix over a larger sound one.

This capture is based on the project’s documentation and release notes, not a hands-on evaluation.