FlowCrew

github.com/cuibuaa/flow-crew · MIT · Node/TypeScript

Give it a brief. Walk away. Come back to a result it has already proven — or an honest “it didn’t work.”

FlowCrew turns a natural-language brief into a supervised crew (planner, coder, researcher, reviewer, QA, supervisor) that plans, executes, retries, and checks its own work unattended for hours. Node/TypeScript, MIT.

What makes it different

• An atom-based execution model. The planner decomposes a brief into a DAG of small, individually verifiable steps. An inner loop converges each stage (gate → fix → retry); an outer loop carries findings and dead ends across runs so a later run pivots instead of repeating.
• Honesty by construction. Before a run is called shipped, an independent re-check has to re-confirm the win. Beat the metric but fail that check and the run is downgraded, not shipped. ceiling (“found nothing”), incomplete (“ran out of budget”) and shipped are distinct, first-class outcomes — never a crash dressed up as a success.
• One engine for research and engineering. The same loop chases a metric (beat a baseline) or satisfies a contract (pass the tests).
• A rehearse “wind tunnel.” flowcrew rehearse <brief> dry-runs the brief through the real scheduler against a scripted agent, in seconds and at zero token cost, to catch engine-contract bugs before spending real compute.
• Human-in-the-loop by default where it matters. A consequential action parks the run and waits for a decision in an approval inbox instead of guessing.
• A live dashboard for run traces, the knowledge graph, reruns, and approvals.

Why I built it

Most agent frameworks are optimized to produce an answer. The failure mode that actually hurts in long autonomous work isn’t a wrong answer — it’s a confident wrong one, discovered hours later. So the design question I cared about was not “how do I get the agent to finish?” but “how do I make the system catch itself when it’s wrong?”