See how they request the agent to use xml to call tools: https://github.com/cline/cline/blob/main/src/core/prompts/sy...
And also: https://github.com/cline/cline/blob/main/src/core/Cline.ts
The way it works is we prompt the LLM to generate code to call the tools directly. We then parse the AST to validate it and restrict it to a very small/safe subset (whitelisted function calls.)
HF smolagents does something similar and there are a few papers out there validating the approach.
But the vendors are training their models specifically on tool use with their expected internal formats so you might as well leverage that if you are starting fresh now.