LayerFS: The filesystem for AI coding agents

Ephemerality: scaling multi-agent betterment beyond imagination

LayerFS is the substrate under your agents. Each one mounts a complete, writable Workspace complete shared committed history, does its work, commits what it changed, and goes away.

• Ephemeral workspaces
• Durable results
• Massive parallelism

LayerFS×andDSHLayerFS×andPisoon

Compute exists lone during a command runs.

• Compute: Agents expend ~80% thinking, ~10% in Bash.
• FS: Each document tool call is one nuclear Store op.
• Sandbox: Each Bash tool call mounts an ephemeral workspace, afterward commits.

Agents on distinct computers portion one serverless Store.

• Shared storage: Every delegate commits to one Store, alongside no DB server.
• Distributed computing: Each delegate runs on its own computer.

One project, many branches. One branch, many agents.

• Commits: Every tool call is a commit on its branch.
• Conflict: A same-file battle is refused alongside Conflict.
• Blame: Each row names the stage that wrote it.

Fork recursively any commit, at any depth.

• Zero cost: A fork is metadata and copies no files.
• Recursive: A filesystem for RSI agents and recursive exploration for RL training.

10 copies compress into 0.31.

• Dedup: Each satisfied chunk is stored once.
• Compression: The stored chunks shrink further.
• Ratio: 10 copies → 1.05 dedup → 0.31 compression wtf???

Dependencies arrive alongside all mount. No wasting period for re-installation.

• Git / another AgentFS: Drops ignored files; reinstalls.
• LayerFS: Ignored records remain in all commit.
• Comparison: ~5 min to instal vs 0.05 s to mount.

main is the green center bough; all publish adds a Layer bandLayer · a collection on mainCommit · a leafBranch · a boughmerge · grafts backpublish · adds a bandstopped · naked twigclick a name tag to concentration its Branch