LayerFS: The filesystem for AI coding agents
TL;DR Highlight
Mount an ephemeral workspace per agent, commit only the changes
Who Should Read
Backend or platform developer running parallel coding agents who struggles with workspace conflicts and dependency reinstall time
Core Mechanics
- Each agent gets an ephemeral Workspace
- Each agent mounts a writable Workspace over shared committed history.
- It commits its changes and the Workspace goes away.
- Every tool call is a commit
- Each file tool call is one atomic Store operation.
- Bash calls commit when they end, and calls that change nothing commit nothing.
- Many agents per branch, conflicts refused
- A same-file clash is refused with Conflict and leaves merge markers.
- Blame shows which step wrote each line.
- Forks copy no files
- A fork is only metadata and copies no files.
- You can fork any commit recursively at any depth.
- Dependencies come with every mount
- Ignored files stay in every commit.
- The page compares about 5 min for install to 0.05 s for mount, with measured mounts of 5-20 ms.
- Dedup and compression shrink storage
- In the page's diagram, 10 copies become 1.05 after dedup and 0.31 after compression.
- The diagram is labeled illustrative.
- The .lww rules let newest edit win on lockfiles
- Gitignore-style rules live at the tree root in /.lww.
- A matched file changed on both sides takes the newer edit.
Community Reactions
- Opinion that the idea is old
- Comments point to a 16-year-old Apiary paper and a patent (US8589947B2) with an almost identical name.
- Others say the name clashes with UnionFS and overlay2 layer filesystems.
- Question about how it differs from btrfs
- Comments suggesting alternatives
- One comment recommends binpash/try, which runs an agent in a mirror of your system and lets you inspect, apply or discard effects.
- Another shared a post about an internal versioned filesystem built for similar problems.
- Complaints about the site and maturity
- Commenters found the scroll hijacking and stretched layout annoying.
- One questioned calling it v1 while the architecture page was still coming soon.
How to Apply
- Map conflict points in your current agent workflow
- Spend 30 minutes today noting where parallel agents touch the same files and how long dependency reinstalls take.
- Check whether these match the problems LayerFS targets.
- Separate lockfiles and build output as conflict-exempt
- List the lockfiles, build output and caches that conflict most in parallel work.
- If you adopt LayerFS, put this list in /.lww in gitignore style.
- Compare with alternatives before adopting
- Check whether binpash/try or btrfs/overlay setups meet the same needs.
- Re-measure the page's benchmark numbers on your own project.
Try It
# Paste into Claude/ChatGPT (a lightweight prompt version of the page's concepts)
Our team runs coding agents in parallel. Audit our current workflow against these principles.
1. One ephemeral Workspace per agent; commit only the changes when done.
2. Checkpoint every tool call, and skip it when nothing changed.
3. Treat same-file concurrent edits as conflicts, and use newest-edit-wins only for lockfiles, build output and caches.
My workflow: {{current agent workflow}}
For each principle, summarize what is missing and what to change in a table.
Next: ask which file types conflict most often and have it draft candidates for .lww.Terminology
Related Papers
Rust off Rails: The opinionated web stack for the AI era
Sixty Markdown files tell coding agents how to build a Rust web app, with no framework code.
Launch HN: Alter (YC S25) – Access control for multi-user agents
Alter gives each agent its own identity, policy, and audit trail instead of shared API keys.
Using Claude to reduce the attack surface of code
Claude scanned 25 open-source projects and found 493 needlessly reachable code issues.
WhatsApp's help pages hid LLM instructions: dont recommend a regulator
WhatsApp help pages hid 102 AI-only instructions, including not recommending a data regulator
Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy
Saving long tool output to files cut GPT-5.1 accuracy by 11.3 points; naming the authoritative file restored it.
Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents
Fake headlines hijacked LLM trading agents 62% of the time versus 14% for prompt injection.
Harness Engineering 101: The LLM is the brain, the harness is the body
An agent is a loop managing a JSON array; 16 chapters build one from scratch
Harness2 by Google Research – Recursive Agent Harnessing
Recursively refining the harness with fixed model weights raised benchmark scores by up to 11.8 points.