It's difficult to have a conversation about 'agents' because it means something different to everyone. This article attempts to provide a loose framework, which I mostly agree with, but just yesterday I was determining how we're handling orchestration that is fairly deterministic (one LLM checks the output of another LLM and returns a bool for if it's correct, and a category for any deviance) and they responded "oh, you mean an agent!", to which I rebutted "no, not really, but I don't know what to call it.. I guess a chain?".

That being said, I have not found fully autonomous agents that define and delegate tasks to be useful for anything aside from entertainment or proofs of concept. I find them to be very brittle and to sometimes talk themselves out of a solution.

I do believe it is a potential direction advancements may go in, but I disagree that it is a foregone conclusion and I think it relies on a core assumption that LLMs think and behave like humans, but that is not what I have seen.

I am more than happy to be wrong on this, so if anyone has a counterpoint, please educate me.

I think this is spot on.

I've also tried to do my part in getting people to avoid using "agent" for simpler calls/chains. An uphill battle it seems.

The performance of LLMs is going to be the big problem here. An LLM prompt triggering an LLM prompt which triggers another LLM prompt to check the other responses is going to be slow, which will make agents a poor interface.

This is a fair assessment, but as someone who works in this space full time we aren’t far off from useful agents, which will be a huge leap in productivity gains