Cloudflare’s new Artifacts tackles a growing bottleneck: agents generate code at a scale ten times larger than human developers can handle. By treating every session, sandbox, or worker as a Git repository, the system lets agents create, fork, and commit without any manual tooling. The result is a distributed, versioned filesystem that feels native to Git.

Artifacts exposes a REST API and a native Workers interface, enabling serverless functions to create repos, generate credentials, and push changes without a Git client. Internally it stores objects in Durable Objects, chunking large files across SQLite rows, and uses R2 snapshots for persistence. The core Git engine runs in a ~100 KB Zig‑compiled WebAssembly module.

Built on Zig, the Git implementation implements SHA‑1, zlib, delta encoding, and the full smart‑HTTP protocol from scratch, keeping the WASM binary under 100 KB. This tight footprint lets Artifacts spawn tens of millions of isolated repositories—each backed by a Durable Object—without exceeding Cloudflare’s per‑namespace limits or increasing latency for developers and teams who build AI.

Within Cloudflare, Artifacts powers internal agents that persist sandbox state, session prompts, and code changes, enabling seamless time‑travel and forked debugging sessions. Because the system mimics Git, developers can instantly share a session URL, let a teammate clone a read‑only fork, and resume work from any checkpoint—cutting onboarding time and reducing merge conflicts for teams.