Service agents versus spending agents
Klarna’s support agent is the textbook production case: 2.3 million conversations in a month, 11 minutes to two, refunds and plan changes inside the ticket. It is also the textbook second-year case: automation of the simple errands, then a reinvestment in human service as the premium path. Read both years.
Spending agents are a different product. Mastercard’s Agent Suite, Agent Pay, and the 9 September 2026 Agent Connect launch are an attempt to standardize discovery → confirmation → consumer-authorized payment so a merchant does not have to rebuild for every model vendor. Klarna’s Agentic Product Protocol is the catalog side of the same problem: 100 million+ products, 400 million prices, 12 markets.
Network rails, not bank apps
Visa’s embeddable assistant is a network product sitting inside issuer apps. Coinbase’s Payments MCP (September 2025) is a tool server for stablecoin movements. Both are “agentic payments” and neither looks like a retail chatbot.
Open Banking Tracker (May 2026) also flags autonomous payment execution and on-chain stablecoin payments as near-term. Near-term is not the same as a named production volume. We log the primitive and wait for the volume.
What a serious payments agent must log
Payer identity, mandate, amount, merchant, MCC, scheme, authorization timestamp, model/prompt version, and the human who can reverse it. If that trail is not queryable, the agent is not a payments system. It is a demo with a card number.