so do we need something like `safe agent execution layer - that is policy enforced` (SEAL) we can manage what should be allowed and what not

agent uses llm to plan the action, but the actual execution happens in SEAL.

any example where it would make sense to start with?

open for thoughts