Cursor details an agent swarm that kept context small and hit 80% on a SQLite-from-docs benchmark

bibryam · x · 2026-07-28

Cursor describes an agent swarm architecture for scaling agentic work without blowing up context or cost.

The system splits work into planner agents and worker agents: smart models break the goal into a tree of subtasks, while cheaper/faster models execute the slices.
Cursor says this design keeps each agent’s context small and lets the swarm handle larger tasks than a single-agent approach.
To validate the design, they reran a hard benchmark: building SQLite from scratch in Rust using only its documentation.
On the same models and time budget, the new swarm outperformed the old one in every configuration. With Grok 4.5, it reached 80% on a held-out SQL test suite in 4 hours, while the earlier swarm degraded and had to be paused before the second hour.
They also varied model assignments across planning and execution; quality stayed similar, but cost changed dramatically.
The post also highlights a custom VCS layer and the broader economics of multi-agent systems.

More from coding & agent