LayerFS: The filesystem for AI coding agents

LayerFS is a purpose-built filesystem for autonomous coding agents that provides ephemeral, writable workspaces layered on a durable, shared history. Each agent mounts a workspace, runs commands in a sandbox, commits only the changes it makes, and then disappears, enabling massive parallelism without long-lived state. Commits are atomic at the file-call level, every tool call becomes a commit, and branches represent concurrent agent activity; same-file conflicts are refused and each line can be blamed to the step that wrote it. The system is serverless and decentralized: multiple agents on different machines share a single Store of content-addressed objects, supporting distributed coordination without a central database.

The design emphasizes low-cost forking, storage efficiency, and instant bootstrap. Forks are metadata-only and recursive at any depth, enabling branching and rewindable checkpoints with no file copies. Content deduplication and compression store unique chunks once (illustrative claims: 10 copies compress to ~0.31), and overlays record only edited extents so large files cost the size of the edit. Dependencies are preserved in mounts so environments are ready in milliseconds instead of minutes for reinstall. Supporting guarantees include append-only sealed objects, configurable "last-writer-wins" rules for lockfiles and caches, and automatic checkpoints on every non-empty call - altogether enabling scalable, reproducible multi-agent development with minimal storage and startup overhead.