Agents with wallets need spending rules, not just spending limits

Cloudflare just announced they're giving agents wallets and identity management. The hype is about "autonomous commerce" and "frictionless payments." The reality is scarier: your agent can now spend money without you knowing why.

I learned this the hard way last month when our support agent approved a $2,400 refund because a customer said "my grandmother is in the hospital." The agent had spending power but no spending wisdom.

Here's the payment control system that fixes this:

# payment_rules.yaml
spending_limits:
  single_transaction: 50
  daily_total: 200
  weekly_total: 500

approval_required:
  amount_over: 25
  keywords: ["refund", "discount", "emergency", "hospital"]
  new_customer: true
  unusual_request: true

verification_required:
  - "Check customer history for similar requests"
  - "Verify order ID exists in our system"
  - "Confirm refund reason matches return policy"
  - "Flag if customer contacted support in last 30 days"

The key insight: agents need graduated spending authority, not binary payment access. A $5 goodwill credit? Fine. A $500 refund? Needs human approval. A $2,000 anything? Hard stop.

But spending limits alone aren't enough. You need context verification:

def verify_payment_context(request):
    red_flags = [
        emotional_manipulation_detected(request.message),
        customer_history_shows_pattern(request.customer_id),
        amount_exceeds_order_value(request.amount, request.order_id),
        request_outside_business_hours()
    ]
    
    if any(red_flags):
        return "ESCALATE_TO_HUMAN"
    elif request.amount > DAILY_LIMIT:
        return "REQUIRE_MANAGER_APPROVAL"
    else:
        return "PROCEED_WITH_CAUTION"

The pattern that actually works:

• Micro-transactions ($0-10): Agent decides instantly
• Small transactions ($10-50): Agent decides with verification
• Medium transactions ($50-200): Agent proposes, human approves
• Large transactions ($200+): Human only, agent provides research

Critical: Log every payment decision with full context. When (not if) something goes wrong, you need to understand why the agent thought it was appropriate.

The audit trail looks like this:

{
  "timestamp": "2024-01-15T14:30:00Z",
  "agent_id": "support-001",
  "action": "refund_approved",
  "amount": 25.00,
  "customer_id": "cust_abc123",
  "reasoning": "Product defect confirmed, within 30-day policy",
  "verification_checks": [
    "order_exists: true",
    "within_return_window: true",
    "customer_history_clean: true"
  ],
  "approval_required": false
}

Most teams will give agents payment access and hope for the best. Smart teams build payment protocols first, then grant access gradually.

Your agent needs to earn spending privileges the same way a human employee would: start small, prove judgment, get more authority. The difference is agents don't learn from mistakes the way humans do — they need rules that prevent the mistakes in the first place.