Gregor Zunic critiques the common use of agent frameworks in machine learning, arguing that the real value lies in the reinforcement learning (RL) model itself, not in the complex abstractions that often accompany these frameworks. He emphasizes that agents are essentially just loops of tool calls and that adding layers of abstraction complicates the process, making it harder to adapt or extend the agents. His experience with early Browser Use agents showed that as soon as changes were needed, the framework became a hindrance instead of a help. Zunic stresses that 99% of the work happens within the model, and effective agent frameworks should allow models to operate with minimal restrictions. He proposes that instead of defining specific actions upfront, developers should give models maximum freedom to perform tasks and only restrict them based on evaluations. This approach led to the development of the BU Agent, which provides raw access to browser controls and APIs, enabling the model to adapt and self-correct when faced with challenges. He introduces the concept of ephemeral messages to manage the vast data generated by browser interactions, preventing the model from losing coherence. Zunic emphasizes that while the mechanics of a simple for-loop are easy, making a robust system requires addressing operational challenges like retries and connection recovery. He concludes with a strong stance against unnecessary abstractions, asserting that simplifying the framework leads to better performance. The insights shared in this article are meant to guide developers in building more effective models without overcomplicating the underlying architecture.