A post citing VC18z asks: Are AI agents a new layer or just a new interface?

A user summarizing a VC18z piece offers a counterpoint to impressive AI-agent demos: at 95% reliability per step, a 30-step job finishes cleanly only about 21% of the time. The post says verification can wipe out the speed advantage, while write access turns an assistant into infrastructure with a bigger attack surface. It also argues that Salesforce, GitHub and Jira already own the workflow—raising the question of whether agents become a new layer or just a new interface on the old one.