Exactly right - the deterministiclayer is the only thing you can actually trust. We landed on the same pattern: LLM handles the understanding, hard rules handl

Exactly right - the deterministiclayer is the only thing you can actually trust. We landed on the same pattern: LLM handles the understanding, hard rules handle the permission. The tricky part is maintaining those rules as the agent evolves. How are you managing rule updates code changes every time or something more dynamic?