LLMs are an algorithm. There will be times it should be used and times it shouldn’t be used because it’s to inefficient and/or to expensive. Then there’s the whole field of decidability… lots of halting problems that have no solution. Who will decide when to make the call that enough tokens have been used and should stop the agent because there’s probably no solution? Who will decide when to use LLM (and which model) and when it makes more sense use another algorithm instead? Deciding which algorithms to use sounds an awful lot like the job of a software engineer.