Comment by cyanydeez

this would be more interesting as a local LLM anlysis; throw out all the costs, and figure out primary-subagent model architecture, and maximize token generation and prefill.

I don't see how anyone can operationalize this information.