SMAshhar - out 2024 I would tend to use o1 models as managers. The agents that I am working on are mostly a diverse set of models. Recently, I built 27 agents AI Architecture for a startup. I used small models to large models, open-source to close-source, depending on the need. For example, if an agent is to go through some text and list down some points, i would use a smaller model, like llama-8B. If its a planner, would opt for llama-70B. For a manager, o1. Used OpenAI as well as Groq wherever necessary. Cost cutting was also a big factor in what i should use.
Fonte: link