A lot of models, even super good ones, fail on tool calls 5-10% of the time, and right now the whole agentic process just stops when that happens. If you just tell the models:
"Please pay attention to developer instructions when formulating tool calls, rethink, and make sure you are formulating correctly."
It resumes the task fine. There should be some sort of automatic tool call failure handling.
A lot of models, even super good ones, fail on tool calls 5-10% of the time, and right now the whole agentic process just stops when that happens. If you just tell the models:
"Please pay attention to developer instructions when formulating tool calls, rethink, and make sure you are formulating correctly."
It resumes the task fine. There should be some sort of automatic tool call failure handling.