Agents is just the LLM prompting itself with subtasks. This can be helpful to compartmentalize larger and more complex tasks. I find the tool calls most important. An agent that can compile and test its own code will make fewer mistakes (assuming the tests weren’t also written by LLMs but even then).

So there’s a self-correcting feedback loop that will improve quality, but is also a lot more expensive. The problem is that the temptation is huge to give more and more complex tasks to the agents and then they go off for 30min or longer and may or may not create complete and utter bullshit because something led them astray or they forgot something one agent did wrong and never fixed it. The end result may work, but it may be less efficient or come with a share of dead code that’s never called etc.

You end up with hundreds of lines to verify. You could just ask an agent to do it, but at some point adding more agents to the pile is not going to help… and probably takes longer than just reading it yourself.