Here’s one that’ll start arguments. MeshXFS is a GPL-2.0 POSIX clustered file system where, per its creator Steve Shoecraft, “no human wrote any of its own code.” Claude Opus, running as an autonomous agent in Claude Code, adapted upstream XFS into a file system many machines can mount read-write at once.
What the AI built
- Two distributed lock managers: one over the network, and one that keeps lock state on the shared disk using SCSI COMPARE AND WRITE.
- Cross-node cache coherency, fencing with SCSI persistent reservations, and survivor-led journal recovery.
- Userspace tools (mkfs, fsck, resize), the test harnesses and release validation.
- The forked xfs/ tree grew from about 226k to about 314k lines.
The starting point is still upstream XFS written by the XFS developers. Everything that makes it clustered came from the agent.
My take
Kernel file systems are about the hardest code there is, so this is a real milestone for coding agents. Would I trust my data to it yet? Not without a lot of time and testing. But the question has shifted from “can it” to “how do we verify it.”
