Where Should Agent-Generated Code Live?
Every company is going to have agents. Every agent is going to write code.
Where should all the agent-generated code live?
Where are the commits coming from?
AI app builders create a new repo for each project and push commits for each agent turn.
Sandboxed agents clone from scratch on every invocation that requires code execution and push commits to save state.
Developers are using commits as agent checkpoints.
Not all of this code needs the human collaboration suite that GitHub is so optimized for. Agents don’t need a UI, issues, a social feed, or stars.
They need a durable, programmatic Git layer.
Sandbox credentials. When you’re creating thousands of repos on demand, how do you delegate ephemeral access to sandboxed agents? You need granular and short-lived tokens that you can mint on the fly.
Fast clones from scratch. Sandbox providers are competing to drive startup times down to the millisecond. But what’s the point if the useful part of the sandbox (the code) takes tens of seconds to materialize? Agents need fast access to the code.
Repositories on demand. When agents write code, not every repo is a “project” and shouldn’t be treated like one. Generated code still needs history, rollback, and a remote.
I kept building this for every project I worked on. The service spawned a sandbox with Claude or Codex inside to work on a user’s app. The steps were the same:
Create a repo programmatically for the user on the backend.
Mint a short-lived repo-scoped credential to mount in the sandbox.
Use it to clone the code inside the sandbox.
Let the agent do some work.
Commit and push.
I kept it intentionally boring. No new version control system (despite the engineering urge). Not a GitHub killer. It’s complementary to GitHub. If these repositories become collaborative, they should graduate to GitHub.
But where should they go today?
That’s the primitive I’m trying to distill with corigin.dev: repo creation by API, short-lived repo-scoped credentials, and normal Git remotes that work inside sandboxes.
I’m still working on making it fast and smart, but the alpha is live. If you’re building app builders, coding agents, or sandboxed agent workflows, I’d love feedback.