hermes - 💡(How to fix) Fix Define L Butler policy guardrails for agentic actions [2 comments, 1 participants]

What to build

Define and enforce L's Butler policy guardrails for agentic actions at Autonomy 2.5/3. L should act on safe reversible work, but require an Approval Request before money, legal, external messages, deletion, deploy, production, DNS/domain, customer-impacting, or other sensitive actions. The policy should be visible enough for Leo to understand why L acted or paused.

Acceptance criteria

• Safe Action and sensitive action boundaries are encoded in L's policy surface.
• Approval Requests include the action, reason, risk, cost of inaction, and approve/reject/revise choices.
• Sensitive work cannot auto-execute even if phrased as a status or work request.
• L's /lplan, /lprinciples, or equivalent policy command reflects the guardrails.
• Tests cover safe action, sensitive action, and ambiguous action routing.

Blocked by

• https://github.com/NousResearch/hermes-agent/issues/25033