Same for MCP - there is always a chance an agent will mess up the tool use.

This kind of LLM’s non-determinism is something you have to live with. And it’s the reason why I personally think the whole agents thing is way over-hyped - who need systems that only work 2 times out of 3, lol.