AI didn't give payments systems a new requirement. It multiplied the one they already had: move the money exactly once.

Visa, Mastercard and American Express have all launched programs for agent-initiated payments, built on protocols like Google's AP2. Those agents are probabilistic. They retry, re-plan, and run around the clock at machine speed. Ledgers can't be.

So rare races become routine. Two agents check the same spend cap and both approve. That's write skew, and snapshot isolation won't stop it, even in databases that label it SERIALIZABLE. And a timeout isn't a failure: the charge may already have landed.

In this session you will learn:

Failover that never pays twice. Consensus replication with zero RPO, replica placement that survives a zone or region, and why a DR test without agents retrying is fiction.
Commit money, mandate and agent state together. No reconciliation window, so a disputed agent payment can be answered months later from one system of record.
Keep fraud reads off the spend path. Follower reads for historical features, fresh reads for velocity and balances, and priority tiers so agent fan-out can't starve settlement.
Enforce limits agents can't talk past. Spend caps enforced by serializable transactions and grants, not app code and not prompts.
Derive idempotency keys that survive agent retries. Keys from the mandate, not the framework, committed with the posting.