Agentic Postcard (Direct Mail)
Use when: an autonomous agent sends a physical postcard or letter to a real address. This skill is a design pattern, not an implementation. Address quality lives in agentic-address.
The one rule
Ink costs money and cannot be unsent. A human approves what prints, and only verified addresses receive it.
Six invariants (do not ship without all six)
- Verify before print. Every address passes verification (see
agentic-address) before entering a batch. Unverifiable ⇒ held for review, never mailed on hope. - Human approves the creative. Front, back, offer, and fine print get a human sign-off per template version. The agent fills fields; it never redesigns the card or invents the offer.
- Spend caps per batch. Per-piece cost × quantity is computed before ordering, against a per-batch and per-campaign cap enforced in code. Over cap ⇒ the batch waits for a human.
- No sensitive content by mail. Credentials, full account numbers, and health or financial details never go on a postcard. Sensitive mail uses a sealed-letter path or no mail at all.
- One send per trigger. Idempotency keys on recipient + campaign + trigger event. Retries and duplicate events never double-mail; reprints need explicit approval.
- Delivery ledger. Every piece appends recipient ref, template version, approver, cost, mail date, and tracking outcome (including returns). Returns feed back into address quality.
Minimal shape
interface MailPiece { to: AddressRef; templateVersion: string; approver: string; idempotencyKey: string }
type MailOutcome = 'queued' | 'held-unverified-address' | 'held-over-cap' | 'mailed' | 'returned'
Verify address → check caps → confirm approval → order with idempotency key → ledger append → track returns.
Anti-patterns
- ❌ Mailing an imported list without verification ("the spreadsheet had zips").
- ❌ Letting the agent write the offer copy unsupervised.
- ❌ No idempotency key, so a retried webhook mails twice.
- ❌ Ignoring returns instead of suppressing bad addresses.