Cursor团队开发的Agent Swarm系统通过分层规划与执行架构实现了多智能体协作，并在复杂编程任务中展现出显著的成本优势与效率提升 [1]。
该系统成功从零开始编写了对应835页SQLite手册的完整Rust代码库 [1]。在使用Grok 4.5模型的测试中，新系统仅需4小时即达到80%的测试通过率，而旧系统在第2小时前因螺旋错误被迫中止 [1]。
不同模型配置的成本差异最为显著 [1]。采用Opus 4.8与Composer 2.5混合模式的总成本仅为1,339美元，而GPT-5.5全栈配置的成本高达10,565美元 [1]。在Fable 5混合模式下，新系统生成的引擎代码仅需9,908行即通过全部测试，相比之下旧系统需要64,305行且效率较低 [1]。
系统通过多项创新机制进一步优化了性能，包括引入第三方仲裁者解决合并冲突、建立Field Guide共享上下文等，将提交速率提升至每秒约1,000次 [1]。这一设计验证了在复杂任务中混合使用前沿模型进行规划与低成本模型执行的策略，能够显著降低AI编程成本并提升稳定性 [1]。
The Cursor team has developed an Agent Swarm system that enables multi-agent collaboration through a hierarchical planning and execution architecture.[1] The system successfully generated a complete SQLite database implementation in Rust from scratch, corresponding to an 835-page manual.[1]
When tested with the Grok 4.5 model, the new system achieved an 80% test pass rate within four hours, whereas the previous system encountered spiral errors and was forced to halt before the two-hour mark.[1] The Agent Swarm architecture incorporates mechanisms such as third-party arbiters to resolve merge conflicts and a shared context system called Field Guide, which increased commit rates to approximately 1,000 per second.[1]
Model configuration choices significantly impact costs. A hybrid approach using Opus 4.8 and Composer 2.5 incurred total costs of $1,339, while a full-stack GPT-5.5 configuration cost $10,565.[1] Under the Fable 5 hybrid configuration, the new system generated engine code comprising just 9,908 lines that passed all tests, compared to 64,305 lines in the previous system, which also demonstrated lower efficiency.[1]
The research demonstrates that for complex tasks, combining cutting-edge models for planning with lower-cost models for execution substantially reduces AI programming costs while improving stability.[1]