DoltLite Beta: SQLite fork with Git-style version control, built with 2k agent PRs
DoltLite 0.50.0, now in Beta, is a fork of SQLite that swaps the B-tree storage for a Prolly Tree-backed chunk store, adding Git-style version control to the familiar SQLite interface. The project was built in just five months by a team of AI agents, using about 2,000 pull requests.
What Is DoltLite?
Everything above the B-tree layer — SQL parser, analyzer, file I/O, and test harness — is stock SQLite. The B-tree layer is replaced with a Prolly Tree (content-addressed B-tree) backed by a single file chunk store. That gives you full version control features: branch, merge, diff, rebase, cherry-pick, reset, and remote operations like push, pull, clone, and fetch from a custom remote or DoltHub.
What Beta Means
- Storage Format Stability: The format is now stable after 12 breaking changes. It's been untouched for 57 releases (3+ months). Future changes will have migration paths.
- SQL Compatibility: Passes 100% of sqllogictest (5.8M queries) and 99.46% of SQLite's 892K TCL acceptance tests. The 4,809 divergences are documented and stem from primary key indexing, chunk-based storage, and no WAL/journal.
- Full Version Control: All core Dolt features are implemented locally and remotely. Works with DoltHub as a sync backend.
- Production Performance: In-memory databases are 10% slower on reads, 60% slower on writes vs SQLite. File-backed are at parity on reads, 10% slower on batched writes. Small autocommit writes are 3.1X slower — use batched writes for performance-sensitive work.
Try It
DoltLite Beta is ready for adoption. Daily performance reports are published on GitHub. If you want Git-like version control for your SQLite databases, DoltLite is worth a look.
📖 Read the full source: HN AI Agents
👀 See Also

ETH Zurich Study: Excessive Context Reduces AI Coding Agent Performance
An ETH Zurich study tested four coding agents on 138 real GitHub tasks and found that LLM-generated context files reduced task success rates by 2-3% while increasing inference costs by 20%. Human-written context only improved success by ~4% with significant cost increases.
Claude Agent SDK Gets Dedicated Monthly Credits for Programmatic Usage Starting June 15
Starting June 15, paid Claude plans receive a separate monthly credit for programmatic usage (Agent SDK, claude-p, Claude Code GitHub Actions, third-party tools). Pro gets $20, Max 5x $100, etc. Usage pauses if credit runs out and additional usage credits are off.

Berkeley Study: All AI Revision Prompts Drift Prose Toward Formality, Even "Preserve Voice"
New paper from Berkeley measures 300 personal narratives through Claude, ChatGPT, and Gemini under three prompt conditions. Every model and condition reduces contractions, first-person pronouns, and narrative closeness — the "preserve voice" prompt only reduces drift magnitude, not direction.

Microsoft Cancels Claude Code Licenses — AI Agents Too Expensive to Scale
Microsoft cancels most Claude Code licenses, shifting engineers to GitHub Copilot CLI. Uber burned its 2026 AI budget in 4 months. Token costs per task may actually increase.