Ephemeral-AI-Lab
Ephemeral-AI-Lab is an open-source AI-agent infrastructure lab focused on making multi-agent software development more parallel, ephemeral, and recoverable. Its flagship LayerFS gives agents disposable writable workspaces over durable shared history, with branching, checkpointing, conflict handling, and deduplicated storage.
At a Glance
- Developers and teams building coding agents
- AI-agent infrastructure and developer-tooling projects
- Research teams running multi-agent exploration or reinforcement-learning workflows
- Users of DeepSeek Harness
- +1 more
AI Tools by Ephemeral-AI-Lab
(1)
LayerFS
Open Source Filesystem for Agents
Discussions
No discussions yet
Be the first to start a discussion about Ephemeral-AI-Lab
Latest News
LayerFS main adds DeepSeek Harness support for Windows Live through a WSL2 worker
LayerFS v1.0.0 released as the first official release
Yifan Xu announced LayerFS v1.0.0 as an open-source recursive filesystem for AI agents
AlphaSignal coverage highlighted LayerFS as a Rust-based filesystem giving AI agents disposable zero-copy forks of a shared base
Products & Services
An MIT-licensed Rust/SQLite content-addressed filesystem for AI-agent workspaces. Agents mount ephemeral writable Workspaces over shared committed history, commit changes as checkpoints, fork recursively without copying files, and branch, rewind, merge, or publish results.
A source-distributed SDK with project, branch, filesystem, and workspace/sandbox APIs for initializing projects, forking branches, reading and editing files, committing, merging, publishing, and mounting workspaces.
A DeepSeek Harness integration that adds LayerFS to agent sessions; users can fork, test, merge, and publish from bash while the DSH web UI exposes project History, Sandboxes, and Files.
Open-source agent sandbox infrastructure for parallel coding agents, providing isolated workspaces, MCP/CLI control, observability, and atomic publication.
Market Position
LayerFS positions itself as an agent-native alternative or complement to Git plus conventional copy-on-write/sandbox approaches: unlike ordinary Git workflows that preserve selected states, it makes every relevant tool-call step a checkpoint and gives each agent a writable workspace over shared history without copying the base. Its published comparisons include Git/other AgentFS approaches, Btrfs plus zstd, libfuse passthrough, kernel overlayfs, and Cloudflare Computer's filesystem benchmark; the differentiation emphasized is recursive zero-copy branching, durable per-step history, conflict-aware coordination, and deduplicated storage.
Leadership
Founders
Yifan Xu
Founder of Ephemeral AI Lab and the GitHub/X account behind the LayerFS releases; describes himself as CS at the University of Toronto, an MSAI dropout at NTU Singapore, and a builder of agent infrastructure.
Executive Team
Yifan Xu
Founder
CS at the University of Toronto; MSAI dropout at NTU Singapore; publicly identifies as a builder of agent infrastructure and is the release author for LayerFS.
Founding Story
The lab's stated focus is 'scaling multi-agent development beyond imagination' through ephemerality. Its initial vision is to let each coding agent mount a complete disposable workspace, preserve useful results as durable shared history, and enable many agents to branch, explore, checkpoint, merge, and recover without copying the base filesystem.
Business Model
Revenue Model
The available sources describe an open-source, source-distributed MIT-licensed project rather than a paid hosted service. LayerFS 1.0.0 is explicitly released as source-only, with no npm package, prebuilt binary, or public image.
Target Markets
- Developers and teams building coding agents
- AI-agent infrastructure and developer-tooling projects
- Research teams running multi-agent exploration or reinforcement-learning workflows
- Users of DeepSeek Harness
- Linux-based development environments and Docker-backed agent sandboxes
- Parallel coding agents working on one project
- Multi-agent software development and coordination
- Agent sandboxing and isolated command execution
- Checkpointing, rollback, and recovery during coding sessions
- Recursive exploration and MCTS-style rollouts
- Reinforcement-learning or agent-trajectory experiments