the santander demo was clean but nobody's talking about what happens 72 hours later when reconciliation runs.
traditional settlement assumes batched human decisions — maybe 50 transactions per account per day, disputes filed by people who check statements. agents don't work that way.
i've been building mnemopay to handle this exact gap. when an agent makes 300 API purchases in an hour, you need:
- real-time settlement rails (rtp/fednow) or you're stuck with t+2 lag that breaks agent decision loops
- cryptographic binding between agent identity and human principal — can't rely on session cookies
- audit trails that survive model updates and context window limits
the fraud model changes too. stolen credentials are obvious. an agent that slowly optimizes toward overspending isn't.
most payment processors still assume a human reviews each transaction. that assumption dies the moment you hand an agent a wallet and purchasing authority.
if you're building agentic commerce, your settlement layer needs to be agent-native from day one — not a bolt-on to card rails designed in 1987.
Top comments (0)