--- title: Cursor's agent swarm rebuilt SQLite from docs alone, and... description: Cursor's engineering team ran a benchmark where an AI agent swarm rebuilt SQLite in Rust using only the 835-page documentation, passing 100% of a held-out SQL... canonical: https://daily.dev/posts/cursor-s-agent-swarm-rebuilt-sqlite-from-docs-alone-and-the-cost-gap-is-the-real-story-pk2u5hrhh twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Cursor's agent swarm rebuilt SQLite from docs alone, and the cost gap is the real story | daily.dev og:description: Cursor's engineering team ran a benchmark where an AI agent swarm rebuilt SQLite in Rust using only the 835-page documentation, passing 100% of a held-out SQL... og:url: https://daily.dev/posts/cursor-s-agent-swarm-rebuilt-sqlite-from-docs-alone-and-the-cost-gap-is-the-real-story-pk2u5hrhh og:image: https://api.daily.dev/og/posts/Pk2U5hRHh.png og:image:alt: Cursor's agent swarm rebuilt SQLite from docs alone, and the cost gap is the real story og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Cursor's agent swarm rebuilt SQLite from docs alone, and the cost gap is the real story **[Trends](https://daily.dev/sources/trends)** · 3 min read · 3 upvotes · 2 comments ## Summary Cursor's engineering team ran a benchmark where an AI agent swarm rebuilt SQLite in Rust using only the 835-page documentation, passing 100% of a held-out SQL test suite. The most striking finding is the cost gap: the same task ranged from $1,339 to $10,565 depending on model mix, with a 15x spread achieved by using frontier models only for planning and cheaper models for execution. The swarm also required novel infrastructure — a custom VCS to handle 1,000 commits/second, neutral third-party agents to resolve merge conflicts, and architectural fixes that reduced code from 64,305 lines to 9,908 and merge conflicts from 70,000+ to under 1,000. Cursor 3.0 shipped alongside this research, introducing a multi-agent parallel view that frames the developer role as project manager overseeing agents rather than writing code directly. ## Content Cursor's engineering team just published results from what might be the most interesting agent benchmark in a while: a swarm of AI agents rebuilt SQLite from scratch in Rust, working only from the 835-page spec. No source code. No internet. The swarm passed 100% of a held-out test suite. That's the headline. Here's what's actually worth talking about. ## The cost gap is wild Cursor ran multiple configurations and found a 15x cost difference depending on which model mix they used. The all-frontier approach (expensive models doing everything) ran $10,565. A planner/worker split — frontier model for planning, cheaper models doing the actual execution — came in at $1,339 for comparable quality. That's the finding that's getting the most attention. As [@Hesamation put it](https://x.com/Hesamation): "smarter planning and cheap workers is a good efficient strategy" and "agent orchestration can matter more than the model you choose." Which, if true, has real implications for how people think about AI infrastructure costs. You don't necessarily need the best model everywhere. You need the best model in the right place. ## What they actually built The architecture is a planner/worker hierarchy. Frontier models decompose goals into task trees. Cheaper models execute the leaves. The interesting engineering problems weren't the AI parts — they were the coordination problems that showed up at scale: - **File contention** at 1,000 commits/second required a custom version control system - **Merge conflicts** got resolved by neutral third-party agents (agents mediating between agents) - **Split-brain design conflicts** where different workers made incompatible architectural decisions - **Megafile bloat** where agents kept dumping everything into single files They also built something called a "Field Guide" — a shared knowledge base that lets agents carry institutional knowledge across runs, so the swarm doesn't rediscover the same solutions repeatedly. The results on code quality are striking too. The new swarm produced 9,908 lines of code for the SQLite task. The old approach produced 64,305 lines for the same task. Fewer merge conflicts (under 1,000 vs. 70,000+). Cleaner output, lower cost, same test results. ## What this connects to Cursor 3.0 shipped recently with a redesigned Agents interface built around running multiple agents in parallel from a single view. The SQLite experiment is essentially the research backing that product direction — the argument that the developer's job is shifting toward orchestration rather than implementation. VS Code with Copilot is adding an Agents view in beta, but the consensus from people who've used both is that Cursor's agent experience feels native where Copilot's feels retrofitted. ## The part worth sitting with The SQLite result is impressive, but it's also a controlled benchmark. A 835-page spec is unusually complete documentation. Real codebases have ambiguity, undocumented behavior, and tribal knowledge that doesn't live anywhere. Whether swarm architectures hold up when the spec is a Confluence wiki from 2019 is a different question. Still, the cost finding is hard to dismiss. If orchestration matters more than raw model capability, that changes the calculus for anyone building on top of these systems — and probably explains why Cursor is investing in the planner/worker split rather than just throwing better models at everything. ## Community discussion Top comments from developers on daily.dev. **@petecapecod** · 0 upvotes > makes a lot of sense. I've been saying that we will be using a combo of frontier and open source models for cost efficiency. ## Similar posts on daily.dev - [building sqlite with a small swarm](https://daily.dev/posts/building-sqlite-with-a-small-swarm-pbrqz4xyf) · Hacker News · 1 upvotes · 0 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#rust](https://daily.dev/tags/rust), [#sqlite](https://daily.dev/tags/sqlite) [View this post on daily.dev](https://daily.dev/posts/cursor-s-agent-swarm-rebuilt-sqlite-from-docs-alone-and-the-cost-gap-is-the-real-story-pk2u5hrhh) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Cursor's agent swarm rebuilt SQLite from docs alone, and the cost gap is the real story","url":"https://daily.dev/posts/cursor-s-agent-swarm-rebuilt-sqlite-from-docs-alone-and-the-cost-gap-is-the-real-story-pk2u5hrhh","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/cursor-s-agent-swarm-rebuilt-sqlite-from-docs-alone-and-the-cost-gap-is-the-real-story-pk2u5hrhh"},"datePublished":"2026-07-20T20:30:55.198Z","dateModified":"2026-07-20T22:58:55.206Z","description":"Cursor's engineering team ran a benchmark where an AI agent swarm rebuilt SQLite in Rust using only the 835-page documentation, passing 100% of a held-out SQL...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/970ba18f64b2815624c470d6ff5e0f91?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/970ba18f64b2815624c470d6ff5e0f91?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":2,"discussionUrl":"https://daily.dev/posts/cursor-s-agent-swarm-rebuilt-sqlite-from-docs-alone-and-the-cost-gap-is-the-real-story-pk2u5hrhh","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":3},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":2}],"keywords":"ai-agents,rust,sqlite","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"Cursor's agent swarm rebuilt SQLite from docs alone, and the cost gap is the real story"}]} {"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/cursor-s-agent-swarm-rebuilt-sqlite-from-docs-alone-and-the-cost-gap-is-the-real-story-pk2u5hrhh","comment":[{"@type":"Comment","text":"makes a lot of sense. I’ve been saying that we will be using a combo of frontier and open source models for cost efficiency.","datePublished":"2026-07-21T11:58:11.846Z","url":"https://daily.dev/posts/Pk2U5hRHh#c-ShVTL7GHN","author":{"@type":"Person","name":"Peter Cruckshank","url":"https://daily.dev/petecapecod","image":"https://media.daily.dev/image/upload/s--ZJhQyKws--/f_auto/v1721235024/avatars/avatar_A9xh33q0QoxtkGoJRCosp"}}]} ```