Cohere 论文：编码智能体团队中便宜的模型只做规划最划算
Cohere 的新论文，实测多智能体写代码怎么省钱：便宜的模型只做规划，强的模型当 lead，成本能砍 70% 掉不了几分。
Cohere 的论文测试了 lead agent 把子任务分发给 worker 的编码智能体团队架构，任务基准为 ProgramBench。多智能体团队胜过单智能体，但成本是其 3 到 7 倍，且会出现重复劳动和额外的代码合并开销。实验显示只负责规划、不写代码的 worker 能达到全阶段参与的同等收益，成本减半。GPT-5.6 Sol 作为 lead 搭配 16 个便宜的 Luna 规划器，成绩距全 Sol 团队仅 1.4 分，成本低 70%。结论是把最强模型放在 lead 位置，用便宜模型只做规划。
New Cohere paper finds that a team of coding agents gets the most from extra agents when they only research and plan, and cheap models can do that job for a strong lead at far lower cost.
A team of coding agents beats 1 agent working alone, but costs 3 to 7 times as much. More agents can also mean duplicate work and extra code to merge.
The researchers built a lead agent that hands subtasks to workers, then varied the workers' role, model and count. On ProgramBench, agents rebuild a whole program from just a working executable and its docs.
Workers that only plan matched the reward of workers used in every stage, at half the cost. A strong GPT-5.6 Sol lead with 16 cheap Luna planners came within 1.4 points of an all-Sol team at 70% lower cost.
On tasks like this, put your strongest model in the lead and use cheap models only for planning, since cheap coders lowered reward.
