SafetyVerified 17 days ago
Separate Draft Autonomy From Send Authority
Let agents prepare freely. Hold anything that publishes, spends, deletes, or messages externally for explicit approval.
The principle
READ and DRAFT can be autonomous. SEND and DESTRUCTIVE must be held. Marketing and coding agents both fail when those classes collapse.
Why it matters
Most AI marketing disasters are not bad writing — they are unapproved publish or spend actions. Governance is cheaper than apology posts.
How to apply it
- Classify tools: read, draft, send, destructive.
- Queue send actions for human approval by default.
- Log who approved and what content shipped.
- Keep kill-switches for automation paths.
Red flags
- Cron jobs that post without a review artifact
- Shared credentials with no approval trail
Quick win
For social ops: write drafts to disk first; publish only after an explicit post instruction.
safetygovernanceagentspublishing