Anthropic rethinks source control for AI-written code
Anthropic published "Redesigning source control at Anthropic." The top comment wants to commit the prompts as the source code and treat AI-generated code as "expensive-to-generate build output." Another commenter compared it to how version control handles huge game and 3D assets. A third was fed up with Git's workarounds for large repos (shallow clones, sparse checkouts, promisor remotes) and just wanted someone to describe the workflow.
This connects to a smaller thread on better ways to track a Git repo's release history, where one commenter noted how little of Git they actually use. Git's model was built for humans typing code slowly. Agents produce changes at a volume and shape it wasn't designed for.
The open question is whether the real artifact is the prompt or the diff. Nobody has a settled answer, and the prompts-as-source idea runs into the fact that model output isn't reproducible.
So what?
Version control for agent-heavy teams is an open product category, and the biggest AI lab is signaling Git isn't enough. If you build dev tooling, look at review, provenance and storage for generated code. If you ship with agents, start recording prompts and model versions alongside commits.