Comment by amadeuswoo

Interesting architecture. Im curious about the workflow when an agent hits a denied action, does it get a structured rejection it can reason about and try an alternative, or does it just fail? Wondering how the feedback loop works between safety kernel and the LLM's planning