> As models are getting better, harnessing will shrink

That can be true only for locally hosted models. The more supplied tools can do, the less data has to be exchanged with OpenAI/Anthropic servers. So bad harness means both higher lag and token usage.