{"slug": "cursor-s-ai-agents-rebuilt-sqlite-from-scratch-and-passed-every-test", "title": "Cursor's AI Agents Rebuilt SQLite From Scratch and Passed Every Test", "summary": "Cursor's AI agent swarm rebuilt SQLite from scratch using only its 835-page manual, passing 100% of a held-out test suite on the first try. The swarm, composed of planner agents on frontier models and worker agents on cheaper models, completed the task in four hours at a cost of $1,339 for the most efficient configuration, compared to $10,565 for a single-model approach. Cursor published the results on April 6, 2026, highlighting a 15x cost reduction, though the company's parent Anysphere was reportedly acquired by SpaceX for $60 billion in June 2026.", "body_md": "*A swarm of AI agents at Cursor rewrote SQLite in Rust using nothing but the database's 835 page manual, no source code, no test suite, and no internet access, then passed every held out test on the first try.*\n\nCursor's research team handed a swarm of AI agents the SQLite documentation and nothing else. No source code. No test suite. No internet connection. Four hours later, the swarm had produced a working reimplementation of one of the most widely deployed pieces of software on Earth, written entirely in Rust, and it passed 100% of a held out SQL test suite the agents never saw during the build.\n\nCursor published the results in a post titled \"Agent swarms and the new model economics,\" written by researcher Wilson Lin and posted to the company's blog on April 6, 2026. It's a striking claim. It resurfaced on Hacker News this week and climbed back up the front page, with developers picking apart what it actually means for how software gets written from here. The short version: SQLite, the database format sitting inside your phone, your browser, and probably three apps you opened this morning, got rebuilt from a manual alone.\n\n## Inside the swarm\n\nThe swarm runs on two tiers of agents. Planner agents, running on frontier models like Opus 4.8, break a broad goal into smaller pieces and hand them off. Worker agents, running on cheaper and faster models like Composer 2.5, execute one narrow slice of that plan and nothing else. A planner never writes code, so its context never fills up with implementation detail. A worker never plans, so it can spend its entire context window on the one function in front of it. That separation is the whole trick.\n\nThe results were fast. Absurdly fast.\n\nCursor's earlier swarm, which spent a week building and running a web browser with no human help, peaked at roughly 1,000 commits an hour and racked up more than 70,000 merge conflicts before the team paused it. The SQLite swarm peaked at roughly 1,000 commits a second. A thousand times faster. And it logged fewer than 1,000 conflicts across its entire four hour run. Standard Git tooling couldn't keep up with that pace, so Cursor built its own version control system to handle it.\n\nBy the four hour mark, the different model configurations were passing between 73% and 85% of the held out test suite. Every one of them eventually reached 100%.\n\n## The catch\n\nQuality converged. Cost did not. Cursor ran the same SQLite build across several model combinations and found the price varied by 15 times depending on which models did the planning and which did the work. A hybrid run pairing Opus 4.8 as planner with Composer 2.5 as worker cost $1,339 total. Running GPT-5.5 alone for the entire job cost $10,565. In the hybrid setup, the pricier planner model produced a small fraction of the total tokens but ate up roughly two thirds of the bill, while the cheap worker model did most of the actual writing for a third of the cost.\n\nHere's the thing: Cursor isn't a neutral observer here. Its parent company, Anysphere, was reportedly the subject of a $60 billion acquisition by SpaceX announced in June 2026, on the back of roughly $4 billion in annualized revenue. Publishing a study showing your own product can cut its cost by 15 times is also a sales pitch. That doesn't make the numbers wrong, but it should make GitHub Copilot's team at Microsoft, Cognition's Devin, and Replit read the fine print closely before repeating the headline.\n\nThere's a real caveat buried in the experiment. SQLite is one of the most exhaustively documented pieces of software in existence, with a specification detailed enough that a patient reader, human or otherwise, could plausibly reconstruct its behavior without ever opening the source. Most enterprise software isn't documented like that. A swarm asked to rebuild an undocumented internal billing system, built on a decade of accumulated logic with no comments and no institutional memory, faces a much harder problem than reading 835 pages and following the rules.\n\nStill, the direction isn't in question. Six months ago, an agent swarm writing production quality code from a spec alone was a research curiosity. Now it's a line item with a dollar figure attached, and that figure keeps shrinking. The next test worth watching isn't whether a swarm can rebuild SQLite. It's whether one can rebuild something nobody bothered to document this well in the first place.\n\n**Also read:** [China Weighs Locking Down Qwen and Doubao Just as the World Adopted Them](https://startupfortune.com/china-weighs-locking-down-qwen-and-doubao-just-as-the-world-adopted-them/) • [Anthropic drops plan to charge extra for Fable 5 as Chinese rivals close in](https://startupfortune.com/anthropic-drops-plan-to-charge-extra-for-fable-5-as-chinese-rivals-close-in/) • [South Korea's Record Exports Show How Deep Its AI Chip Bet Really Goes](https://startupfortune.com/south-koreas-record-exports-show-how-deep-its-ai-chip-bet-really-goes/)", "url": "https://wpnews.pro/news/cursor-s-ai-agents-rebuilt-sqlite-from-scratch-and-passed-every-test", "canonical_source": "https://startupfortune.com/cursors-ai-agents-rebuilt-sqlite-from-scratch-and-passed-every-test/", "published_at": "2026-07-21 05:04:37+00:00", "updated_at": "2026-07-21 05:11:00.545639+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-agents", "ai-research", "ai-tools", "developer-tools"], "entities": ["Cursor", "SQLite", "Wilson Lin", "Anysphere", "SpaceX", "Microsoft", "GitHub Copilot", "Cognition"], "alternates": {"html": "https://wpnews.pro/news/cursor-s-ai-agents-rebuilt-sqlite-from-scratch-and-passed-every-test", "markdown": "https://wpnews.pro/news/cursor-s-ai-agents-rebuilt-sqlite-from-scratch-and-passed-every-test.md", "text": "https://wpnews.pro/news/cursor-s-ai-agents-rebuilt-sqlite-from-scratch-and-passed-every-test.txt", "jsonld": "https://wpnews.pro/news/cursor-s-ai-agents-rebuilt-sqlite-from-scratch-and-passed-every-test.jsonld"}}