Cursor’s agents rebuilt SQLite from 835 pages and passed every test

Signals Inbox·July 20, 2026·AI DevTools

Cursor gave a swarm of coding agents SQLite’s 835-page manual, but no source code, tests, executable, or internet access. The agents rebuilt a Rust database that eventually passed every held-out test. The bigger surprise is economic: similar results cost anywhere from $1,339 to $10,565 depending on which models planned the work and which models wrote the code.

Q1What did Cursor actually do?

In its official research post, Cursor says it asked a swarm of agents to rebuild SQLite in Rust from its 835-page manual. The agents could not see SQLite’s source code, test suites, executable, or the internet. Every new swarm configuration eventually passed 100% of Cursor’s held-out SQL test suite.

Q2Why is SQLite a serious test?

SQLite is not a weekend app. A database engine has to parse SQL, plan queries, store data, handle transactions, and reproduce lots of small edge cases. One wrong behavior can break real software. Giving agents documentation instead of source code tests whether they can understand a complete system, design it, and make its behavior match.

Q3Is passing every test unusual?

Yes. In ProgramBench, a May 2026 study covering 200 full-program reconstruction tasks, no tested model completely solved even one task. A June benchmark called MirrorCode found that its strongest model averaged 56% across 25 programs. Cursor’s result is not directly comparable because the harness, compute budget, and tests differ, but 100% still makes this a serious long-horizon coding result.

Q4What changed from Cursor’s older swarm?

The old swarm mostly drowned in its own activity. One Grok run produced 68,000 commits in two hours and accumulated more than 70,000 merge conflicts. The new version produced fewer than 1,000 conflicts over four hours. Cursor separated agents into planners that make decisions and workers that handle smaller tasks, then gave them shared design notes and stronger automated review.

Q5Why did the cost vary so much?

Cursor tested several model mixes. Using GPT-5.5 for both planning and implementation cost $10,565. Using Opus 4.8 for planning and Composer 2.5 for implementation cost $1,339. Both eventually passed every test. In the cheaper setup, the entire worker fleet cost only $411, suggesting that expensive intelligence may be needed for a few key decisions, not for every line of code.

Q6Did the agents rebuild the real SQLite?

Not exactly. They created a Rust replica that matched Cursor’s held-out SQL tests. Passing those tests does not prove complete compatibility with every SQLite feature, extension, performance requirement, security property, or production edge case. This is a strong behavioral reconstruction experiment, not a drop-in replacement ready to run inside millions of devices.

Q7So what is the real signal?

The scarce part of software development may be moving upward. Cursor’s agents handled thousands of implementation decisions once they received a detailed specification and a working management system. If this holds on real projects, companies may spend less time writing every feature and more time describing exactly what the software should do. The winners may be the teams with the clearest specs, tests, and architecture, not simply the biggest model budgets.