LayerFS mounts an agent workspace in 5 to 20 milliseconds

LayerFS is an MIT-licensed filesystem for AI coding agents, published this week by Ephemeral AI Lab. Its page at layerfs.ai sets out one job: give each agent tool call a complete, writable workspace, in a few milliseconds, over a history every agent in a fleet already shares. The Show HN thread for it had 69 points when picked up.

The step it removes

A coding agent that runs a tool, reads the result, and runs another tool spends most of its wall-clock time setting up the working directory for each step. npm install on a working repository takes around five minutes. Running fifty agents against the same repository, each needing its own clean workspace, is fifty npm installs in a loop.

LayerFS reports a mount in 5 to 20 milliseconds. The workspace is an overlay: a thin writable layer on top of a shared history, so dependencies the agent did not change are not copied. The project page states that ten copies of the same workspace compress to 1.05 times its deduplicated size on write and down to 0.31 times after a canonical pass. A single example shown on the page records 48.0 MB of logical content stored as 31.2 MB on disk.

Fork, checkpoint, converge

Two patterns from the page the project is designed for. First, zero-cost recursive forking at any commit depth: an agent can take a copy of the current state, run a step, decide against it, and throw the fork away without rewriting the whole tree. Second, automatic checkpointing with conflict detection: many agents writing in parallel each take their own fork, and the server reports which files overlap when they merge.

Coordination is described as serverless, with agents reaching the storage directly rather than through a central scheduler. The licence is MIT and the release line reads v1.0.0.

Where it would matter in a working team

Most of the wall-clock time a fleet of coding agents burns goes to restoring, modifying and tearing down the working directory for each tool call. Model inference is a smaller fraction of the total. A team using Claude Code, Codex or Cursor in parallel on one repository today sees this cost in wall-clock time and in a bigger cloud bill at the end of the month. If LayerFS holds the mount time it claims on your own repository, the step stops dominating the budget.

Two things to check before adopting it on a real team. First, the compression figures are reported for the project's own example and have not been measured on an outside production repository, so run it against a copy of yours before counting on the number. Second, Ephemeral AI Lab is a small maintainer group, so the usual open-source questions apply around release cadence and security response time. The code is on GitHub for anyone to fork and freeze.

Source

LayerFS, from Ephemeral AI Lab, with the project source at github.com/ephemeral-ai-lab.