Loom: A Git-Based Coordination Layer for Multiple AI Coding Agents

Running several AI coding agents against the same repository exposes a gap in git: branches and commits carry no notion of intent, so agents overwrite each other's work and conflicts only surface at merge or CI time, after tokens and hours have already been spent reconciling divergent rewrites.

Loom addresses this by sitting on top of git rather than replacing it. Every task runs in its own git worktree for physical isolation. Before editing, an agent declares an 'intent lease' — a goal plus file-glob scope — which warns other agents of overlap at declaration time rather than blocking them, letting them self-partition work via a single MCP call. The shared mainline ('fabric') only advances through a green verification run plus a consent gate, and performs an honest file-level merge: if a file changed in both the mainline and a thread, the landing is refused with a clear 'fabric moved under you — rebase' message instead of silently overwriting either side.

Dead or abandoned sessions become orphaned threads — complete with goal, acceptance criteria, and a recent checkpoint — that any agent can adopt and resume. Loom integrates with MCP clients like Claude Code and supports serverless syncing between collaborators via hidden git refs on a shared repository. It's explicitly not meant for redundant agents racing the same task, nor as a replacement for git history or normal human PR workflows — its value scales with the number of concurrent writers and edit frequency, making it aimed squarely at fleets of coding agents.