Agent harnesses have barely been available for a year, and reasonably reliable for only half, and there's already fatigue. I think this says less about whether LLM's will actually be able to program and more about how mentally exhausting AI-assisted programming can be, which involves a higher frequency of decision making and reading an astronomical amount of both code and prose if you actually want to stay on top of what the agent is doing to your codebase. This personal/psychological exhaustion and negative sentiment is now being inaccurately transferred into a pessimistic prognosis for the advancement of the technology itself.