**Title

We've been running 3-4 Claude Code agents in parallel on the same repo for the last month, and the biggest time sink wasn't LLM tokens — it was agents silently undoing each other's work or building conflicting approaches to the same module. That's why we shipped an open-source coordination layer that sits between your agents and git, and honestly it's cut our rework by maybe 60%.

Here's how it works in practice:

Detection hooks — Before any agent writes code, a lightweight scanner checks pending changes across all active sessions. If two agents touch the same function signature, file, or even similar naming patterns, it raises a flag with a diff preview so you can merge intentions rather than code.

Design conflict resolution — Built a small consensus protocol over JSON manifests. Each agent writes a plan.json in its scratch space describing what files it intends to modify and the interfaces it depends on. The coordinator diffs these against each other and against the main branch, warning before any hard conflicts land.

.coordignore — Just like .gitignore, but for agent overlap. You declare regions of the repo each agent "owns" so they don't step on each other. Works especially well for monorepo setups where frontend and backend agents share the same tree.

The whole thing is ~800 lines of Python + a couple of git hooks, and it runs in under 200ms per check. We open-sourced it under MIT if anyone wants to try it on their multi-agent Claude Code or Cursor workflows.

Real-world impact: before this, we'd regularly find half-finished refactors that two agents had started independently with incompatible assumptions. Now those get caught in pre-commit, and we resolve them in a 5-minute huddle instead of a 2-hour git surgery session.

If you're running parallel AI coding agents, I'd argue coordination shouldn't be an afterthought — it's the difference between agents amplifying each other or fighting over the same namespace.