How much logic should we actually keep outside the LLM?

been thinking about this while working with LLM-based agents.

We still use retries, fallbacks, state machines and routing logic to make agents more reliable. But with newer models getting better at tool calling and handling failures, I wonder how much of this we really need to maintain ourselves.

Personally, I'd still keep permissions and critical actions deterministic. But for things like deciding the next step or recovering from a failed tool call, I'm not sure we need to hardcode everything.

Curious how others are handling this in production. Have you actually removed orchestration logic because newer models handle it better or has the complexity just moved elsewhere?