Source · CMS snapshot (seed) — this guide body is served fromcontent/generated/cms-snapshot.json, not the code catalog alone. Staff path: Content Studio (:3030/experience/sites) → publish guide →MCV_CONSOLE_URL=http://localhost:3030 pnpm --filter @mcv-tech/marketing-shell content:seed→ rebuild:3050. Offline:MCV_CMS_SNAPSHOT_FILE=… pnpm content:seed.
The outcome you are walking toward
A written policy — one page is enough — that classifies every action your agents may take, names who releases each gate, and lets an auditor reconstruct any automated decision without interviewing your team.
1. Inventory actions, not features
List what agents would actually do in your operation: adjust a roster, issue a refund, message a member, move money. Features are marketing; actions are what policy governs.
2. Classify by consequence
Every action lands in one of two classes: reversible work agents may do inside declared boundaries, and irreversible work that sits behind a named human gate. The classification scheme and boundary model are defined in governed agent boundaries — adopt it rather than inventing your own vocabulary.
3. Name the gatekeepers
A gate without a named owner is a suggestion. For each irreversible action class, record who may release the gate and under what conditions. Approval gates for autonomous agents covers how releases are recorded and scoped per brand.
4. Turn automation on last
Only after the classification and gate list exist do you enable the automation itself. The reasoning — boundaries before autonomy — is argued in Agents need boundaries before autonomy. Every gate decision lands in the audit trail from the first action, which is what makes the policy enforceable rather than aspirational.