Letting an LLM write one small program that calls your tools in a loop beats round-tripping every call through the model. Here is how I built that for UC AI with Oracle MLE, and why the sandbox was the hard part.
