Hello everyone, has anyone else encountered issues with GPT-5 series models not following instructions when using the official ReAct strategy? The models often directly output process data such as Observation and Action, and sometimes, even after generating an Action, they skip tool invocation and directly output the result.
Below are some example screenshots:
By examining the Agent’s delve log, the desired format should have tool invocations placed within action_input, rather than inside thought.
Could this issue be related to the internal implementation (internal Prompt) of the Agent Strategy plugin? Is it less compatible with GPT-5 series models?