The Missing Abstraction for AI Agents: The Agent Filesystem (turso.tech)

Turso-backed AgentFS SDK introduces an "agent filesystem" — a POSIX-like abstraction that stores an AI agent’s entire runtime state inside a single SQLite file. Designed for sandboxed and autonomous systems, AgentFS unifies files, key-value state, and tool-call audit logs so every file created, state change, and external tool invocation is captured, queryable with SQL, and portable as one snapshotable file. That unified approach improves observability, reproducibility, and trust: you can copy a DB to reproduce an exact execution, run SQL for debugging, and maintain an immutable audit trail for postmortems or compliance. Technically, AgentFS implements a filesystem schema on SQLite/Turso with separate inode and dentry tables (inodes hold file contents/metadata; dentries map paths and speed lookups), a JSON-serialized key-value table for structured state, and an append-only toolcall table for immutable audit logs. The SDK exposes familiar APIs (JS/TS and Rust) — e.g., AgentFS.open('./agent.db'), agent.fs.writeFile(...), agent.kv.set/get(...), and agent.tools.record(...) — letting agents capture rich runtime telemetry automatically. By replacing ad-hoc mixes of logs, object stores, and local filesystems with one queryable DB, AgentFS simplifies agent development, makes debugging systematic, and scales observability in production agent sandboxes.