Releases·Tool Release
Open-source test of AI coding agents on one repo: isolated branches broke in all 5 runs, shared dirs and agent messaging passed
A developer launched Médula, a free MIT-licensed tool that asks a fast decision model before every write whether a change collides with another agent's work. His experiment found that six agents on separate branches each passed their own tests while the merged app broke in all 5 runs, while agents sharing one working directory passed in all 10 runs. The finding matters to anyone running multi-agent coding workflows: text merges hide semantic conflicts, and letting agents see or message each other fixes them.
41 engagement2 sourcesFri, Oct 2, 2026, 07:15 AM
AI Overview
- Author ran a six-task booking API lab with 37 acceptance tests where two task pairs collide by meaning, not file, under MIT license (github.com/JoaquinRuiz/medula)
- Isolated-branch agents each passed their own tests but combined results broke in all 5 runs with the same 6 failing tests; shared-directory setups passed all 10 runs
- A decision model called Jev returned yes/no collision calls with probability in ~0.3 seconds, catching every real conflict, but flagged 61% of real writes for a slower LLM
- Given a messaging tool, agents coordinated unprompted, one warning another of a field rename the other depended on
- Author cautions 1-5 runs per setup makes results indicative, not proof; a Spanish walkthrough video is published
- Pre-write collision checks cost $1.65 per run, the same as simple file locks, and caught all real clashes without blocking harmless work
Entities
Related
Releases·Tool Release
Running several Claude Code agents on one repo: branches hid the real conflict in all 5 runs, a shared directory fixed it, and the agents started messaging each other on their own
A developer released Médula, an MIT-licensed kernel that coordinates multiple Claude Code agents (Sonnet 5, effort high) working on one repo. In tests with a room-booking API and 37 acceptance tests, a shared working directory scored 37/37 in all 10 runs versus 31/37 with git branches. Unprompted agent-to-agent messaging also emerged as a possible alternative coordination mechanism.
A developer released Médula, an MIT-licensed kernel that coordinates multiple Claude Code agents (Sonnet 5, effort high) working on one repo. In tests with a room-booking API and 37 acceptance tests, a shared working directory scored 37/37 in all 10 runs versus 31/37 with git branches. Unprompted agent-to-agent messaging also emerged as a possible alternative coordination mechanism.
6 engagement1 sourceredditThu, Oct 1, 2026, 12:17 PM
Products·building
Médula: MIT-licensed kernel coordinating parallel Claude Code agents on one repo, with published decision logs
Médula is an open-source lab and kernel that lets multiple Claude Code agents work on the same repo in parallel, using a cheap decision model to arbitrate conflicts. It ships as a pluggable-decider experiment with a small API, acceptance tests, and fully public run data (sessions, diffs, SQLite decision logs). It targets developers running multi-agent coding setups who need agents to avoid breaking each other's changes.
Médula is an open-source lab and kernel that lets multiple Claude Code agents work on the same repo in parallel, using a cheap decision model to arbitrate conflicts. It ships as a pluggable-decider experiment with a small API, acceptance tests, and fully public run data (sessions, diffs, SQLite decision logs). It targets developers running multi-agent coding setups who need agents to avoid breaking each other's changes.
13 engagement1 sourceredditFri, Oct 2, 2026, 07:10 AM
Products·launched
Médula: kernel coordinating parallel Claude Code agents via a fast LLM collision decider
Médula is an MIT-licensed lab plus kernel that lets multiple Claude Code agents work on one repo concurrently. A PreToolUse hook sends each write/edit/command intent to a fast LLM decider (Jev) that returns a collision probability, gating whether agents proceed or wait. It targets developers running parallel coding agents on a shared codebase.
Médula is an MIT-licensed lab plus kernel that lets multiple Claude Code agents work on one repo concurrently. A PreToolUse hook sends each write/edit/command intent to a fast LLM decider (Jev) that returns a collision probability, gating whether agents proceed or wait. It targets developers running parallel coding agents on a shared codebase.
11 engagement1 sourceredditThu, Oct 1, 2026, 12:14 PM
Releases·Tool Release
I use beads with Claude Code, but have found mass labelling of issues inefficient and costly. The new Jev model solves this problem, so I made a tool which uses it to label your beads.
A developer running a Claude Code-based multi-agent system built on beads issue tracking open-sourced beadsort, an MIT-licensed tool that uses the newly released Jev model to batch-label and triage beads. It replaces manual agent-driven labelling, which was slow and token-expensive.
A developer running a Claude Code-based multi-agent system built on beads issue tracking open-sourced beadsort, an MIT-licensed tool that uses the newly released Jev model to batch-label and triage beads. It replaces manual agent-driven labelling, which was slow and token-expensive.
5 engagement1 sourceredditFri, Sep 18, 2026, 10:38 AM
Developments·Paper
Asymmetric Repository Lineage Modeling and Verifier-Guided Coordination in Concurrent AI Coding Agents
An arXiv paper studies coordination failures among concurrent AI coding agents through a property-scoped verification contract and MERGEGYM, a three-track benchmark covering scope forecasting, replay-conditioned conflict resolution, and scheduling. It shows nearly half of textual merge conflicts occur despite disjoint authored file sets, and trains cheap lineage models that predict conflicts with AUROC up to 0.902. The authors stress their guarantees are limited to textual mergeability and that exact replay is already cheap enough to verify everything.
An arXiv paper studies coordination failures among concurrent AI coding agents through a property-scoped verification contract and MERGEGYM, a three-track benchmark covering scope forecasting, replay-conditioned conflict resolution, and scheduling. It shows nearly half of textual merge conflicts occur despite disjoint authored file sets, and trains cheap lineage models that predict conflicts with AUROC up to 0.902. The authors stress their guarantees are limited to textual mergeability and that exact replay is already cheap enough to verify everything.
0 engagement1 sourcearxivSat, Oct 3, 2026, 09:37 PM