Cursor launches AI agent swarm for ultra‑fast code commits

Cursor has built an AI‑driven agent swarm that separates responsibilities between planner agents and worker agents. Planner agents employ powerful frontier models to recursively decompose a coding goal into a tree of sub‑tasks, while worker agents use faster, cheaper models to execute those tasks. This role split addresses a context‑management problem that causes single‑agent drift on long jobs.

The new swarm can process about 1,000 commits per second, a ten‑fold increase over its earlier system that handled roughly 1,000 commits per hour. To support this speed, Cursor created a custom version‑control system that mitigates failure modes unseen in human teams. The architecture includes a "split‑brain" design where planners may unintentionally duplicate work, and a neutral agent that steps in to resolve merge conflicts. Decisions are logged in shared design documents that code references at compile time, and bloated files are flagged for external agents to split into smaller modules. Multiple review perspectives and a self‑maintained "field guide" improve reliability, allowing agents to patch code outside their primary area when needed.