Current systems for AI agents are inefficient and often lack accountability. They confuse the outputs generated by large language models (LLMs) with unstructured data, focusing too heavily on human interaction through vision rather than text. The author argues that while we can build systems that treat LLMs as key components, this approach is not sustainable. As LLMs evolve, they will match human capabilities, turning these systems from essential tools into optional ones. A significant challenge lies in the gap between the speed of technological advancement and the accountability for the systems we create. While LLMs have drastically accelerated software development, our ability to track and manage these systems has not improved correspondingly. This creates what the author calls an "accountability sink," where failures occur without a clear human owner to diagnose and correct them. The lack of understanding about how systems function leaves organizations vulnerable to breakdowns and inefficiencies. The author highlights that the fundamental issues stem from inadequate infrastructure for system observation and understanding. Many systems are built on fragile foundations, relying on luck rather than reliable processes. As the reliance on LLMs grows, it's essential to develop agent-native systems that prioritize human oversight. These systems should be "radically observable" and "radically deterministic," allowing users to clearly see how systems operate and ensuring consistent outputs. The call to action is clear: focus on creating systems that enhance human understanding and ownership rather than merely catering to AI agents.