Human confirmation is required when the next action changes scope, authority, data exposure, or recovery risk. Examples include:
- Connecting or disconnecting a production site or user.
- Changing billing, plans, account ownership, or product entitlements.
- Publishing, deleting, restoring, or bulk-changing production content or data.
- Sending form, social, email, webhook, or API traffic to a real destination.
- Uploading private evidence or granting credentials and persistent access.
- Merging, releasing, or deploying code beyond the explicitly authorized repository and environment.
Low-risk local edits and read-only research can proceed when they are clearly within scope. The agent should stop when a missing choice would materially change the result, explain the consequence, and request the smallest decision needed.