LayerFS: The filesystem for AI coding agents

What it claims

LayerFS is a versioned, serverless filesystem built around the tool calls of AI coding agents. Its cost argument is that compute should exist only while a command runs. The page says agents spend about 80% of their time thinking and about 10% in Bash, so a sandbox held open for a whole session sits mostly idle. Storage is the persistent layer, and compute attaches to it per call.

How it works

The page names two operation types with different handling:

1. A file tool call (read, write, edit) is one atomic operation against the Store.
2. A Bash tool call mounts an ephemeral workspace, runs, then commits the result.

Each tool call becomes a commit on its branch. One project holds many branches, and one branch can take many agents at once. A same-file clash is refused with Conflict rather than merged. Blame is tracked per line, and each line names the agent step that wrote it.

Forks are metadata only and can be taken from any commit at any depth, recursively. The page pitches this at recursive self-improvement (RSI) agents and at recursive exploration for RL training.

Storage stores each content chunk once and then compresses the stored chunks. Ignored files, such as installed dependencies, are kept in every commit. The page contrasts this with Git and "other AgentFS" tools, which drop ignored files and force a reinstall.

Evidence

The page gives one storage figure: 10 copies of a workspace take 1.05 units after deduplication and 0.31 after compression, relative to one copy. It does not say what the workspace contained or how it was measured. The mount comparison is about 5 minutes to reinstall dependencies against 0.05 s to mount, with no project, package manager or hardware stated.

Limits

The page is a feature list with vendor figures. It has no documentation, install commands, API shape, pricing, licence or supported platforms. It does not say how Conflict surfaces to the agent or how a refused write is retried. It does not describe the Store's backing service or its consistency guarantees across machines. The 80%/10% time split and the 0.31 ratio come without a source workload. Keeping ignored files in every commit also means build artifacts and caches enter history, and the page gives no figure for their storage cost beyond the deduplication ratio.