L7 · Agents
Guardrails and Permission Models
Classify actions by reversibility and require approval for the risky ones.
Four of the five stages only read and write text, so they cannot break anything by themselves. Every guardrail you write belongs on the fifth.