Cursor's latest agent swarm system demonstrates that cheaper models can efficiently handle most coding tasks when paired with frontier models for planning. The system was tested by asking both the new and old swarms to rebuild SQLite in Rust using only documentation, without source code or internet access. Every configuration of the new system reached 100 percent on the test suite, while the old swarm struggled with merge conflicts. This marks a significant shift for Cursor, as agent fleets have transitioned from a research project to a core product. With Cursor 3, developers can now run entire fleets of AI agents in parallel, enhancing productivity and efficiency. Source: thedecoder
The system divides agents into two roles: planners using powerful frontier models to break down goals into smaller tasks and workers using faster, cheaper models to execute those tasks. This division helps address the context problem faced by single agents, which must keep track of both the goal and current task simultaneously, leading to drift during long jobs. Cursor claims that swarms scale less through parallel work than by splitting context between planners and workers. The new system also reached 1,000 commits per second, prompting Cursor to develop its own version control system to manage the high throughput. Source: thedecoder
Cursor tested several configurations, including GPT-5.5 solo, Grok 4.5 solo, Opus 4.8 as planner with Composer 2.5 as worker, and Fable 5 as planner with Composer 2.5 as worker. The new system outperformed the old one in every configuration, achieving scores between 73 and 85 percent after four hours, while the old runs scored between 11 and 77 percent. Every configuration of the new system later reached 100 percent. The old system produced 68,000 commits in two hours, mostly wasted work, compared to the new system's 1,000 commits per second. Source: thedecoder
