The Bitter Lesson of Agent Frameworks

Complex agent frameworks often fail because their abstractions constrain the inherent capabilities of powerful, pre-trained models. A more effective approach is to build minimal agents, treating them as a simple for-loop of tool calls with an explicit completion signal. This philosophy suggests providing the model with a maximal action space, such as raw browser controls, and then restricting it based on evaluation, rather than starting with limited, hand-crafted actions. This 'inversion' of control allows the agent to leverage the model's full reasoning ability and adapt to new tasks without fighting a rigid framework. The less code built around the model, the more effective the agent can be.