Function calling vs React Agent #36
Description
Previously, Windows-Use users experienced issues with the ReAct agent, as it was often unstable in terms of providing responses due to a lack of adherence to the correct response format.
To address this, we decided to implement a functional calling approach while also including the thought and evaluation as the tool parameters to enhance stability and taking to account the non reasoning models.
Additionally, the reason the model takes time to perform actions is primarily due to the generation of "thinking" tokens before the actual output tokens. I noticed a significant difference in speed when I disabled the reasoning_effort setting to "none."