Cursor's Multi-Agent System Rebuilds SQLite in Rust, Slashing Costs 15x

The Cursor team conducted a long-horizon coding test where a group of multi-agents successfully rebuilt SQLite in Rust using only an 835-page manual without providing executable reference code. The rebuilt code passed 100% of the held-out test suite. This experiment not only validates the feasibility of using agents for complex engineering tasks but also reveals the massive cost impact of model combinations.

Key Details and Cost Differences

The core conclusion of this experiment is that in complex agent tasks, having frontier models handle planning, orchestration, and critical decision-making, while delegating most execution to cheaper models, can significantly reduce total costs. According to data shared by @ZainHasan6, the total cost difference between various model combinations is drastic: using Opus 4.8 + Composer 2.5 costs $1,339, while Grok 4.5 costs $1,928. The maximum cost difference between various approaches reaches $20,057, representing a 15x cost span.

Swarm Architecture and Test Design

According to @AI寒武纪, the Cursor team refers to this multi-model agent system as an 'agent swarm' architecture. Its design core is to use a strong model (like GPT-5.5) as the planner for task decomposition and decision-making, and weaker models (like Composer 2.5) as workers for execution, thereby drastically optimizing costs. Furthermore, @HamelHusain points out that the key to this test lies not just in model capabilities, but in demonstrating how to maximize AI agent effectiveness through excellent test set design and control over held-out evaluations.

2026-07-21 ~ 2026-07-22 · 8 related posts

Primary sources

1 near-duplicate retellings: Sam_Witteveen