Open lab: does a cheap decision model keep parallel coding agents from colliding?
jokiruiz · reddit · 2026-10-02
The author open-sourced Médula (MIT): a kernel coordinating multiple Claude Code agents on one repo, plus the full experiment dataset. The test bed: a small API with 6 tasks and 37 acceptance tests, with two pairs of tasks colliding by meaning, not file.
- One branch per task: git let the real conflict through and the same 6 tests failed in all 5 runs, even though every agent finished green
- Shared directory: all 10 runs passed, with per-file locks or the kernel; the kernel blocks only real collisions
- Fast path uses a hosted decision model, unsure on 61% of real write requests, escalating to a slower LLM
- Decider is pluggable: any model outputting a collision probability fits, with a 100-pair calibration set
- All data public: sessions, diffs, SQLite logs of every decision with probability, latency and cost
Author seeks blind human labels, calibration pairs from real runs, and new scenarios (schema migration, dependency bump). Caveats: 1-5 runs per mode, thresholds fitted on measured pairs.
Related event: Médula Open-Sources Coordination Kernel for Parallel Claude Code Agents(2 posts)→
More from coding & agent
- mitsuhiko becomes Earendil's "chief MCPO" and teases hot takes on MCP elicitation — mitsuhiko · 2026-10-02
- ProVer paper: LLM judge picks key trajectory steps, beats GRPO by up to 9.9% — omarsar0 · 2026-10-02
- A 3:25 MV made end-to-end by Opus 5.5 in Claude Code, skill open-sourced on GitHub — FinanceYF5 · 2026-10-02
- Local Cursor over thousands of markdown files works solo, but team sharing stays unsolved — Petr275 · 2026-10-02
- Open-source Open-Dots launches right after OpenAI's Dots, self-hostable and works with any LLM — FinanceYF5 · 2026-10-02
- Simple CLAUDE.md tweak makes Claude Code output far easier to scan — julianharris · 2026-10-02