was able to get most reliable results from some combo of: - in prompts ensuring it's clearly explained which tools are available and when to use them, you'll have to figure out how the llm prefers them being referred to - in the tool config making the names and descriptions semantically clear - using an llm that's good at both voice and tool calling (i like 4.1-mini) - splitting the call flow into nodes and providing only the tools required for the specific task being handled