« All posts

How Replit's Snapshot Engine Makes AI Coding Agents Safe

Replit's Bottomless Storage Infrastructure uses copy-on-write and git-based versioning to make AI coding agents safe, reversible, and sandboxed.

Replit has built a storage and compute system that lets AI coding agents modify code and databases without risking irreversible damage. Its Bottomless Storage Infrastructure, introduced in 2023, exposes virtual block devices backed by Google Cloud Storage, split into immutable 16 MiB chunks referenced by manifests. Because chunks never change, copying a filesystem is just a manifest copy—constant-time regardless of size—enabling both fast project remixing and point-in-time checkpoints via copy-on-write.

Code changes are tracked with standard git commits created at each Agent checkpoint, backed by an immutable, append-only git remote that can recover history even if the entire filesystem is deleted. Databases get the same treatment: an unmodified PostgreSQL instance runs on the same storage layer, making it versioned and forkable, while a strict dev/production split ensures the Agent only ever touches a development database.

For engineers, this means AI agents can experiment more freely—adding logs, testing risky changes, or spawning multiple parallel attempts at a problem—because every change is cheaply reversible. Replit's roadmap points toward transactional compute and parallel agent simulations, aiming to make AI-assisted development both safer and more exploratory.