The problem

It seems that many frontier models get into a failure mode where they call the same tool over and over. Including gemini-flash-2.5 and gpt-5. This is typically a symptom of a model that is too small but these models seems to route to a small model and get stuck into this loop.

Of course there is times when calling the exact same tool more than once may be correct and necessary. We need to allow the model to provide a nonce and reasoning comment for calling the tool again to allow it to be called. As long as the reasoning is unique within the current AgentRun we should let it through. (parameters _nonce and _duplicate_reasoning)