For a complex AI use-case, there are currently two schools of thought...

School 1: You need to spend a lot of time and thinking (and investment) in creating a chain of specialist agents that each do a different job.

School 2: Just hit the smartest, best chain-of-reasoning model with an amazing one-shot prompt and let it do the job.

I think, wherever possible, you want to attempt to go with path 2.

Why?

Because the first approach gives you fine-grained control and observability.

However, the second approach allows me to bet that...

a) These billion-dollar frontier model companies will create smarter and smarter models that just get the job done without babysitting.

b) A model with full context of what it's done and what it's planning to do next is going to do a better job than different agents trying to infer intent and picking up the ball from each other

What do you think?