多模型混合调用的问题在于,高级模型容易被低级模型的错误结果误导,且缺乏有效的校验机制,导致最终结果偏差不断积累。而在这方面,Cursor的做法更为稳健:只要不启用 auto 模式,模型的智能水平就能保持一致。同时,它通过内置的linter和apply model(Cursor 自研模型)帮助agent快速发现并纠正错误
@mrabino1@benhylak In cursor at least it goes wild, ignores prompts instructions, project rules, and will add tons of additional ideas or unasked for features
It will sometimes rewrite entire files, create new ones instead of using old ones, cheat at passing tests etc it really depends