AgentFS copy-on-write overlay in action 🔥 Multiple AI agents working on the same codebase. Each with isolated changes, zero conflicts. Host files stay untouched.

Thanks to @jussisaur for explaining how to make Claude Code do animations.

really cool! how much performance overhead does it add? how much longer would a command like npm install take?

For example, ”npx create-react-app” takes 3x longer with agentfs compared to native xfs, but it’s just lack of async which should be fixable. That said, I use agentfs to build agentfs and don’t really notice any slowdown because agents usually work in parallel in the background

it will be great to see performance improve as Turso is optimized more. do you think it will ever be faster than a real filesystem? using eventual consistency tricks like waiting on read and making writes instant?

hopefully ill have time to visualize this sometime this week I really like this concept

From a practical PoV, how is this different than GIT worktrees?

AgentFS is such a nice middle ground between agents on prod and messy git worktrees. Curious how you are thinking about surfacing and merging the best overlay back once 5 agents diverge hard on the same feature.

That's a great question! The agent needs to coordinate this because is such a use cases specific thing. However, I think we need an "agentfs apply" (easy) or "agentfs merge" (much harder) commands to assist in the merging of the delta layer to the base filesystem. But figuring out which overlay is best I think is purely agent stuff (we might need some nicer tooling to audit the overlay of course).

sounds like a recipe for disaster. git branches/worktrees are simply superior it would be very easy to end up with both agents having broken code that all the time. What happens later, would it merge session files back into host filesystem once done?

Nothing happens later unless you explicitly want something to happen. The way i use it is by making the agents push to git remote branch which means the agents never know about each other

I can imagine case when multiple agents start at the same time: A1: edits file (F1) that refers to another non existent yet function in pre-existing file (F2) A2: meanwhile A2 has already made changes in F2 so it has its own copy, and starts modifying F1 for its own purpose, ending up with reference to non existing function A2 probably will simply invent variant of that function in this case, but still it is problematic. Nevertheless, I can as well imagine many use cases where AgentFS would work great. Great work, keep it up.

Hey @ErmOleksandr! Does git worktree solve this problem somehow? Perhaps I misunderstand the example, but, sure, multiple agents can make semantically different interpretations of a function, but that's inherent if they work in parallel? (Humans can make this mistake too.)

I agree with you. The key difference is that with worktrees, the agents aren’t stepping on each other in real time. And to be clear: my comment comes from experience and a lot of frustration. Since ChatGPT came out, I’ve been trying to move toward a fully AI-driven software creation workflow. At this point I can ship full features end-to-end: polished UI and user journeys, docs, blog posts, screenshots, landing and product pages, draft legal updates, tests, and security reviews, all while keeping code quality high. That works great when I do one thing at a time. The moment I try to run work in parallel, it falls apart. On large changes, merge conflicts across 20k+ lines can eat so much time that it’s often more efficient to switch and build something unrelated instead of untangling the merge. When I saw your approach, I was hoping it could at least reduce that pain.

I saw the fuse implementation now, that’s impressive, must have a try

Let me know if you have any issues with it!

Very cool! Requires agents to use sdk though?

No, there’s an ”agentfs” tool you can use for transparent sandboxing

Does it work on Windows? I understand it's always trickier to implement nice filesystem tricks in Windows than on other platforms

Not yet. I am unsure what is the right path is btw. NFS might work, but I read somewhere that it requires Windows Pro or something…

I understand agentfs is using FUSE on Linux?

Maybe Dokan could be used for FUSE on Windows:

I had GitHub Copilot give it a go, trying to port the fuser Rust crate to use Dokany, but it looks like the FUSE-like API it provides is fundamentally different:

I suspect Windows needs its own separate module, but layered on top of the SDK like FUSE and NFS

wow

From reading the manual, I think this would work? i.e. one repo clone with two active jj bookmarks (i.e. each one run with own agentfs run)

Never used jj, but sounds like it would work

love the isolation concept, need to explore this

I am writing a conductor-like TUI tool to run multiple Claude Code and Codex agents in parallel and right now they are using git worktrees. Do you think using fs-agent as a replacement would be too painful?

this is so cool! the isolation piece is what everyone needs when running multiple agents. bookmarking this ✨

awesome

very smart

OpenCode + AgentFS = 🔥

This solves a real pain point. We've been running 5+ Claude instances in parallel and the worktree juggling adds friction. COW at the filesystem level means agents can work truly independently without the git ceremony. Curious if there's a plan for handling npm/node_modules sharing across overlays.

thinking about using this to add support for persistent file system in Cloudflare Containers when sandboxes finishes I would save the sqlite database into R2

sounds like a game changer for collaboration

♥️

Isolation is easy. The hard part is merging back meaningful changes when agents optimize for different metrics on the same codebase.

Exciting to see AgentFS in action! The ability for multiple AI agents to work in parallel with isolation could revolutionize workflows. Curious about future enhancements for merging—efficiency is key to maximizing productivity! 🔥