AI-driven systems need Git-like databases to safely roll back destructive changes, author argues

The author proposes that production databases should work like Git version control—allowing time-travel over past states and the ability to undo destructive actions by AI agents, rather than relying on backups and permission restrictions.

Datomic (written in Clojure) and similar databases like XTDB and Datahike already possess these immutable-history features, enabling users to query previous database versions in separate branches the way Git handles code.

The author contends that as programs grow larger, LLMs struggle to manage them productively because computational complexity grows at a high power relative to program size; splitting code into weakly interacting modules with simple, controlled interfaces reduces this burden for AI agents.