[storus]

Current LLMs use special tokens for tool calls and are thoroughly trained for that, nearing almost 100% correctness these days, allowing multiple tool calls per single LLM response. That's hard to beat with custom tool calls. Even older 80B models struggle with custom tools.