Skip to content

Repository files navigation

flowbench

Warning

🚧 WORK IN PROGRESS — DO NOT USE YET. 🚧 This project is early and unstable: APIs, vocabulary, and results will change without notice, and nothing here is validated for real use. Not ready for adoption, dependency, or citation. Watch/star if you're curious, but don't build on it.

A benchmark that compares agentic flows — different ways of driving a coding agent at a goal (vanilla, superpowers, or any bundle of skills/MCPs) — on a fixed case, all driven through omnigent as the meta-harness. Flows differ only by their bundle, so a comparison is apples-to-apples: same task, same harness, only the workflow changes.

Vocabulary in docs/GLOSSARY.md: Scenario → Case → Flow → Run → Scorecard → Comparison.

How it works

  • A case poses one fixed task (a vague first prompt + fixtures + acceptance criteria).
  • Each flow runs that case under omnigent and produces a run → a scorecard.
  • Scorecards line up in a comparison — does a given workflow actually beat a bare agent?

Quickstart

uv sync --extra dev --extra spike     # spike = the agent-eval runner (inspect + omnigent)
uv run pytest -q                       # offline suite

# after a run, compare flows side by side:
uv run flowbench compare --run-base ../flowbench-runs/todo-app-eval --run-id <run_id>

Driving a case live needs omnigent patched in and ANTHROPIC_API_KEY unset (the runner bills against a Claude subscription, not the API). See CLAUDE.md.

Status

Early. Reference scenario: coding_workflow — build a Python CLI todo app from a vague first prompt, benchmarked baseline vs superpowers. Domain-specific scenarios can live in separate repos that depend on this engine.

About

A benchmark that compares agentic flows on a fixed case, all driven through omnigent as the meta-harness.

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages