Do you mean not just LLMs, but agents? Is this jot avoided by narrowing your scope and just using the chat interface that also may not produce what you're hoping for, but at least can't muck about in your existing code?