SQLite is the perfect database for AI agents—but nobody knows how to use it. Many agent developers I've talked to make the same point: SQLite is ideal for capturing agent state, handling auditing, ensuring compliance, and managing snapshots. It's lightweight, all data represented in one single file. But when I ask them what the interface should look like, nobody can really answer. What should the schema look like? How do we make it work across different agent frameworks? How do we handle the filesystem transparently? After some iteration and conversations with folks building agents, I've identified the three missing pieces: 1. A standardized schema specification. We need a common way to represent agent state, actions, and outcomes. This isn't just about storing data—it's about enabling an ecosystem of tooling. When everyone uses the same schema, we can build debugging tools, replay systems, and compliance analyzers. 2. A programmatic SDK. Raw SQL obviously isn't enough. Developers need type-safe APIs, automatic migrations, and framework integrations. The SDK should feel native whether you're using a framework or rolling your own agent system. 3. A transparent filesystem layer. People building agent compute infrastructure actually want a transparent agent filesystem that works with bash, curl, and friends. Agents shouldn't have to think about where their data lives. They just read and write files, and our system handles the rest. That why me and @glcst at @tursodatabase have been hacking on AgentFS (github.com/penberg/agentfs)—a specification and implementation for agent state management. If you're building agents and fighting with state management, I'd love to hear from you. And if you want to help shape this standard, check out the repo and join the discussion.