MySQL Time Travel: A Safety Net for AI Agents in Production

AI agents are coworkers now. They write your migrations. They generate your tests. They commit to your repos. They deploy your services. They write and execute SQL against your production databases. When humans did the work, mistakes were caught in review. When agents do the work, mistakes are caught in production.

Every major Database has a way to ask “what did this row look like 20 minutes ago?” Everyone except MySQL. Oracle has Flashback. SQL Server has temporal tables. PostgreSQL has extensions.

And MySQL? MySQL has point-in-time recovery from backups, which means hours of RTO, downtime, and FK cascade reconstruction by hand. When agents are writing SQL at machine speed, that gap stops being acceptable.

This talk is about the layer MySQL never had: an indexed, queryable history of every row change, accessible from wherever you already work. AS OF queries through ProxySQL. MCP for AI assistants. REST for incident runbooks. A web dashboard for forensic exploration.

We’ll cover the binlog mechanics that make it possible, the FK cascade reconstruction problem nobody solves correctly, and end with a live demo: an AI agent investigates a real production incident in real time, through multiple interfaces, with a human reviewing and applying the recovery SQL.