the filesystem for AI coding brokers
Compute exists solely whereas a command runs.
- Compute: Agents spend ~80% considering, ~10% in Bash.
- FS: Each file software name is one atomic Store op.
- Sandbox: Each Bash software name mounts an ephemeral workspace, then commits.
Agents on completely different computer systems share one serverless Store.
- Shared storage: Every agent commits to at least one Store, with no DB server.
- Distributed computing: Each agent runs by itself laptop.
One undertaking, many branches. One department, many brokers.
- Commits: Every software name is a commit on its department.
- Conflict: A same-file conflict is refused with Conflict.
- Blame: Each line names the step that wrote it.
Fork recursively any commit, at any depth.
- Zero price: A fork is metadata and copies no recordsdata.
- Recursive: A filesystem for RSI brokers and recursive exploration for RL coaching.
10 copies compress into 0.31.
- Dedup: Each content material chunk is saved as soon as.
- Compression: The saved chunks shrink additional.
- Ratio: 10 copies → 1.05 dedup → 0.31 compression wtf???
Dependencies include each mount. No losing time for re-installation.
- Git / different AgentFS: Drops ignored recordsdata; reinstalls.
- LayerFS: Ignored recordsdata keep in each commit.
- Comparison: ~5 min to put in vs 0.05 s to mount.
