GitHub Was NOT Made for AI Agents (So Cloudflare Built Their Own)

Highlights

• Artifacts provides a Git-compatible, distributed file system built specifically for AI agents, allowing programmatic management of thousands of repositories.

• Cloudflare implemented a Git server by compiling a Zig-based Git implementation to WebAssembly, which runs on durable objects.

• Worker bindings for Artifacts reduce network round trips compared to using standard REST APIs.

• Developers can fork repositories in parallel within durable objects to automate tasks, refactorings, or PR reviews.

• Artifacts data persists in a SQLite database on the durable object, ensuring availability even if the service restarts.

• Unlike S3 files or other file systems, Artifacts natively supports the Git protocol, making it uniquely suited for agent-based workflows.

Timeline

1. 00:00 - 00:42

   Limitations of GitHub for Agents

   Standard platforms like GitHub prioritize human-centric features such as followers and discussions, which are unnecessary for AI agents. Agents require high-volume, programmatic control over repositories to perform tasks like parallel code reviews and automated refactoring. A distributed file system designed for agents must bridge this gap by prioritizing Git compatibility over social features.

   • GitHub design focuses on human social interactions rather than agent requirements.
   • Agents require programmatic repository creation, forking, and deletion at high scale.
   • Git operations are native to agent training data, making Git compatibility a primary requirement for agent-friendly tools.
2. 00:43 - 02:01

   Architecture of Cloudflare Artifacts

   The technical foundation of Artifacts relies on a custom Git server implementation written in Zig, compiled to Wasm, and deployed on durable objects. This architecture allows any client, including workers using isomorphic Git or standard HTTP clients, to interface with the repositories. The use of worker bindings streamlines communication, increasing performance compared to standard REST API implementations.

   • Cloudflare built a Git server using a Zig implementation compiled to WebAssembly (Wasm).
   • The system runs on durable objects to store repository data.
   • Worker bindings allow for more performant interactions by reducing REST API round trips.
3. 02:02 - 04:28

   Implementing Agentic Workflows

   Developers can create complex workflows by forking a baseline repository into multiple artifacts, each assigned to an individual agent task. Because the worker bindings do not currently expose all Git commands, agents often utilize in-memory storage and isomorphic Git for temporary file manipulations before committing and pushing changes. All operations persist in the SQLite database attached to the durable object, allowing for recovery if the object restarts.

   • Artifacts allow for forking repositories to run parallel tasks in isolated environments.
   • Agents interact with repositories by cloning, reading, writing, and committing changes via the Git protocol or in-memory file systems.
   • State persistence is guaranteed by the underlying SQLite database within the durable object.
4. 04:29 - 06:01

   Potential Extensions and Future Possibilities

   The potential applications for Artifacts extend to complex, multi-agent systems where orchestrators coordinate code changes across many repositories. Integration with additional Cloudflare tools, such as sandboxes for execution or browser instances for UI verification, allows agents to test and validate their work. Native Git compatibility remains the core differentiator, making it highly effective for agent-managed source control.

   • Orchestrator agents can manage multiple task-specific agents and merge changes into a main repository.
   • Cloudflare sandboxes can execute code in non-JavaScript languages or run shell commands for validation.
   • Artifacts distinguish themselves from storage solutions like S3, ZeroFS, or JuiceFS through native Git compatibility.