July 21, 2026AgentsCoding

Cursor Built SQLite From Scratch With an Agent Swarm

Cursor's new post on agent swarms is the most concrete look yet at what happens when you stop running one coding agent and start running a thousand. The setup is two roles arranged as a tree. Planner agents, powered by the smartest models, take a goal and split it into pieces, delegating downward. Worker agents, running cheaper and faster models, actually write the code. The whole thing peaks at around a thousand commits per second.

The benchmark that tells the story: rebuild SQLite from scratch, in Rust, using nothing but its documentation. The old swarm spiraled and had to be paused before it hit the two-hour mark. The new one, running Grok 4.5, reached 80 percent in four hours. Same task, wildly different outcome, and the only thing that changed was how the agents were organized and which models sat in which seat.

That last detail is the real lesson, and it's an economics lesson dressed up as engineering. You don't put your most expensive model everywhere. You put it where the thinking happens, the planners doing decomposition, and you flood the cheap fast models into the grunt work. Intelligence becomes a resource you allocate, not a dial you crank to max. The team's earlier stunt, a swarm building a whole web browser for a week with no human, proved the ceiling exists; this post is about making the floor useful. Read it at cursor.com/blog.
← Previous
MCP's Biggest Rewrite Yet Ditches Sticky Sessions
Next β†’
Microsoft Turns YouTube Tutorials Into Agent Skills
← Back to all articles

Comments

Loading...
>_