Currently in ToolCallingAgent, memory_step.model_output is augmented by appending a formatted string representing every tool call.

• I think it is not necessary. It was originally intended to better visualize tool calls in Gradio interfaces. But these interfaces now visualize the content of step.tool_calls directly
• it has the unintended effect of showing LLMs a wrong history of past messages. The artifact that it created for me is that the model would see tool call descriptions in history, thus it would start to write itself Tool call {id}: calling {name} with {arguments} instead of using JSON action format.

Thus we can remove it!