Skip to main content

​
Policy Rules

Autopilot levels decide how much the agent may execute on its own. Policy rules are the complementary control on what it may do — a guardrail you write, evaluated in the actor gate before any action executes.

​
Effects

Each rule has one effect (strongest wins when several match):
  • Block — the action is never executed. It’s marked rejected with the policy reason and shows up in the Actions queue.
  • Require approval — the action is never auto-approved, even at a high autopilot level. It lands in the review queue for you to approve manually.
  • Log — allowed through, just recorded.

​
Matching

A rule matches an action when both conditions hold:
  • Action types — one of the selected types (e.g. asc_review_reply, revenuecat_experiment). Empty = any type.
  • Keywords — at least one keyword appears in the action’s title/rationale (case-insensitive). Empty = any.
Matching is deterministic — no LLM, so it’s predictable and cheap.

​
Setup

Settings → Policy rules. Add a rule, pick an effect, scope it to action types and/or keywords. Examples:
  • “Never auto-reply to reviews” → Block, type asc_review_reply
  • “Pricing changes need my sign-off” → Require approval, keywords price, paywall, trial
  • “Track every metadata change” → Log, type asc_metadata_update

​
API

Rules live on preferences:

​
See also

  • Modes — agent mode + autopilot levels
  • Actions API — the queue where blocked/gated actions surface