DoltLite hits beta: Git-style version control for SQLite, built by agent PRs

DoltLite, a fork of SQLite that adds Git-style version control, has reached beta with version 0.50.0. According to DoltHub's announcement, which surfaced on Hacker News's front page, the project arrived roughly five months after launch and took about 2,000 pull requests to get there — pull requests largely produced by a team of AI agents.

DoltHub says DoltLite began as a test of Gas Town, an agent orchestrator built by Steve Yegge. The company had wanted an embedded version of Dolt for years but had ruled out rewriting Dolt's storage engine in C or Rust by hand. SQLite offered an existing host for that engine, and the agent experiment supplied the labor.

What the fork changes

DoltLite leaves everything above SQLite's B-tree layer untouched: the SQL parser and analyzer, the filesystem layer and the test harness are stock SQLite. Below that line, the B-tree layer is replaced by a Prolly Tree — a B-tree whose nodes are addressed by their content — backed by a single-file chunk store. DoltHub uses the same structure across its other products to power version control.

Because the query layer is plain SQLite, DoltHub avoided the hardest part of its usual work, building a SQL engine, and only had to write a version-controlled storage layer in C. The payoff for users is branch, merge and diff operations on SQLite databases, plus a conflict-aware sync engine built on push, pull, clone and fetch against a custom remote or DoltHub. The release also ships with the Dolt Workbench GUI, including an agent mode, and DoltHub points to dolt_reset('--hard') as the escape hatch when an agent damages the data.

What beta means

DoltHub defines the beta label with four claims: a stable storage format, SQL compatibility, full version control, and production-level performance.

On format stability, development required twelve backward-incompatible storage format changes, each forcing users to stay on one version or manually dump and reimport databases. The current format has now held for 57 releases, more than three months, and DoltHub says any future breaking change will come with a supported migration path.

On compatibility, DoltLite passes all of sqllogictest, a suite of 5.8 million queries — expected, since the query layer is SQLite's own. It passes 99.46% of SQLite's 892,277 TCL-based acceptance tests, with 4,809 documented divergences, each carrying a listed reason. The main causes: tables are keyed by primary key rather than rowid to support version control, storage uses chunks rather than pages, and there is no WAL or journal sidecar, so those tests are skipped.

Version control coverage includes local operations — branches, merges, diffs, rebases, cherry-picks and resets — alongside the remote sync operations described above.

Performance

Reads are close to SQLite parity; writes carry a version-control tax. DoltHub publishes a nightly benchmark report on GitHub comparing the two on a sysbench-style workload. The latest numbers show in-memory DoltLite databases 10% slower on reads and 60% slower on writes, while file-backed databases match SQLite on reads and are 10% slower on batched writes.

The outlier is small autocommit writes, which run 3.1x slower: roughly 125 microseconds in SQLite versus about 400 microseconds in DoltLite on a small GitHub runner. DoltHub's advice is to lean on batched writes in performance-sensitive workflows, noting that even the slow path stays sub-millisecond per write.

Why it matters

For data tooling, this puts reviewable history, branching and conflict-aware sync into SQLite, the default embedded database for countless applications. That enables auditable schema and data changes, safe experimentation on branches, and a hard reset for agent-driven edits — a meaningful safety property now that agents increasingly touch production data.

For software engineering, the release is a data point on agent-driven development at scale. DoltHub states that a coordinated agent team wrote a C storage engine — work the company had shelved for years as too costly to attempt by hand — in about 2,000 pull requests. Field reports on functionality and stability cited by DoltHub are positive so far. The obvious caveats: the quality and performance figures come from the vendor's own testing and benchmarks, and a beta label still implies rough edges. DoltHub is now inviting users to try DoltLite for embedded use cases and to raise issues on its Discord.