DoltLite Beta: SQLite Fork Built by 2,000 Agent PRs
DoltLite went Beta yesterday — v0.50.0, five months after launch. The interesting number isn't in the changelog: about 2,000 pull requests, most of them written by AI agents running on Steve Yegge's Gas Town orchestrator. DoltHub's CEO started the project to test whether an agent team could pull off a real engineering problem. Verdict: a SQLite fork that does Git's job.
What shipped in DoltLite Beta
DoltLite is a fork of SQLite where everything above the B-tree — SQL parser, analyzer, file-layer, test harness — is stock SQLite. The B-tree itself is swapped for a content-addressed Prolly Tree backed by a single-file chunk store. What you get is Git semantics on a .db file:
- Branch, merge, diff — version control for databases, not just schema migrations.
- Push, pull, clone, fetch — conflict-aware sync, with DoltHub usable as the remote.
- Beta stability — storage format frozen after 12 breaking changes (stable for 57 releases), full SQL compatibility, and it passes SQLite's own test battery plus custom Dolt oracle tests.
Why 2,000 agent PRs matter
DoltHub wanted an embedded Dolt for years; rewriting the storage engine in C or Rust was always a bridge too far. A team of agents crossed it in five months. If the most conservative category in software — a database fork that must survive SQLite's entire test suite — ships on agent labor, "agents write code" stops being a demo and becomes capacity planning.
The question flips from can agents build this? to who reviews 2,000 PRs? That's an org-chart problem, not a code problem. Same pattern Dispatch keeps flagging: agent output is becoming infrastructure, and the bottleneck is moving to review, licensing, and trust — not capability.
The verdict
Run the .db, sure. But read the PR history first — that's the paper trail on what autonomous software development actually looks like at production scale.