Skip to content

Repository files navigation

agent-tool-benchmark: hands-on benchmarks of AI agent tooling. Same machine, same tasks, full scripts.

Release License

Every completed benchmark here is a real run of the real tools: within a comparison, every tool runs the same task suite on one machine in one session. Each one ships with its full methodology, rerunnable scripts, fixtures, and honest limitations, so you can reproduce the numbers instead of trusting a vendor chart. The search-API comparison is still pending (harness committed, no keys on the machine it was written on), so it publishes no numbers.

Benchmarks

Category Comparison Date Verdict
browser ego lite vs browser-use CLI 2026-07-28 ego lite wins for local everyday agent tasks (cross-origin iframe piercing, parallel Spaces, zero tab interference); browser-use CLI remains the only option for servers, CI, and cross-platform
editing udiff vs SEARCH/REPLACE vs str_replace vs apply_patch vs Cline vs whole-file 2026-08-19 git apply / GNU patch remain the only appliers with zero silent corruptions across 13 scenarios; Cline's fallback ladder and whole-file rewrite tie for most corruptions (4); three appliers silently append a trailing newline; OpenAI apply_patch safely rejects overlapping hunks but corrupts tab-indented files
documents pymupdf4llm vs markitdown vs docling vs pdftotext 2026-08-19 pymupdf4llm wins outright: 9/10 headings, correct two-column order, byte-exact CJK, ~1 s/doc at ~120 MB; docling matches structure but needs ~12 s/doc warm (39 s on its first cold call) + ~6 GB and silently maps CJK ideographs to lookalike radicals; markitdown/pdftotext emit plain text (0 headings)
retrieval ripgrep vs ctags vs tree-sitter vs jedi (find-all-references) 2026-08-19 jedi: perfect precision, follows re-exports and separates same-named methods, but blind to getattr dispatch; ripgrep: highest recall (0.95) including string-based dispatch, half its hits are noise; ctags indexes definitions, not references (0.00 recall); an aliased constructor call defeated all four tools
retrieval zg (zvec-grep) hybrid vs BM25 vs ripgrep (NL intent → code) 2026-09-07 On 12 natural-language intent queries over requests v2.32.3, zg answered 11–12/12 vs ripgrep's 6/12 — the vocab-gap tier is binary (zg 3–4/4, rg 0/4); but zg costs ~360 tokens and ~1.6 s per query vs rg's ~46 tokens at 5 ms, and pure BM25 --fts matched hybrid, so the official tool-call-reduction claim rests on avoided agent round-trips, which this does not measure
search Exa vs Tavily vs Serper vs Brave vs Firecrawl vs Jina pending Harness + query suite committed; awaiting API keys for a captured run — no numbers published until then

Principles

  • Same machine, same tasks. Within a comparison, every tool runs the identical task suite on the identical hardware in the same session — except where a tool structurally cannot attempt a task (browser-use has no local parallel mode, so browser task 05 is ego-only and scored n/a).
  • Reproducible. All task scripts, generators, fixtures, and committed results are in the repo. The browser suite reruns per task with bash tasks/<tool>/<task>.sh; the other suites rerun with their run.py, which writes the committed results.json. Cloned or fetched inputs (the requests checkout, the rg/ctags binaries) are pinned and re-downloaded by fetch_corpus.sh / fetch_tools.sh.
  • Honest about limits. Single runs, network variance, and untested modes (such as autonomous LLM planning) are stated, not hidden.
  • No invented numbers. Every figure in a results table comes from a captured run. Where the capture is committed in machine-readable form it is the suite's results.json; where it is not (the browser suite's single-run charts), the suite says so. Vendor claims are labeled as claims. Figures whose source is outside this repo — a vendor's own published claim, an upstream checkout, a fetched release asset, a live endpoint probe — are checked against a dated capture in docs/snapshots/, which records the source URL and the fetch date; those captures are point-in-time and their freshness is not verified by the gate.

Planned

  • Agent terminal and sandbox execution environments
  • Spreadsheet agent tools
  • Search-API captured run (harness ready, awaiting keys)

License

MIT

About

Hands-on, reproducible benchmarks of AI agent tooling — same machine, same tasks, full scripts

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

2 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages