I assume the question is rather which LLM can cover most of the tasks while delivering decent quality. I would prefer an architecture using different LLM for different tasks rather like 'specialists' instead of simple 'agents'. I used to take the main task and divide it in smaller tasks and see what can I use to solve the problem. Sometimes rule-based approaches can be already enough for a sub-task and LLM would be not only overkill but also more difficult to implement and maintain.