Cursor test shows cheap coding models suffice if a frontier model plans

Cursor had its upgraded agent swarm, which separates planning from execution, rebuild SQLite in Rust using only documentation and no source code or internet access. Every configuration of the new planner/worker split scored 100% on the test suite, while the prior swarm design failed on merge conflicts it created itself.

Why it matters: It's an evidence point for a specific cost-cutting architecture in agentic coding: let one strong model plan and route work, and cheaper models can execute reliably. That pairs with other items in this space, like Kuaishou's KAT-Coder-V2.5 training on verified coding environments and Poolside's small open-weight coding model beating larger rivals, suggesting the coding-agent race is shifting toward orchestration efficiency rather than just bigger single models.