Make consequential actions explicit and reversible

Mark irreversible or costly actions, support dry-runs and idempotency keys, and require confirmation before they execute.

Why

Missing confirmation on consequential actions is how agents wipe data or overspend. The canonical AX framework (agentexperience.ax — Agent Interactivity Patterns) establishes that sensitive flows — purchases, credential changes, destructive actions — are where digital services need explicit hooks to maintain user trust. Anthropic identifies human-in-the-loop checkpoints for irreversible actions as a core production pattern. The 12-Factor Agents framework calls this Factor 6 ("Design Pause Points") and explicitly labels it "not optional." Cross-links to the human anti-pattern AX-HUM-05 (agentic sludge).

Do

Flag side effects and offer a dry-run.

Don't

Let a destructive call run on first inference.

Artifact

POST /v1/refunds HTTP/1.1
Idempotency-Key: 5f3c...-once-per-intent
X-Dry-Run: true
Content-Type: application/json

{ "order_id": "ord_123", "amount": 4200 }

{ "dry_run": true, "would_refund": 4200, "currency": "usd", "reversible": false,
  "confirm_with": { "header": "X-Dry-Run", "set_to": "false" } }