Partner · Blockchain Life 2026 — Dubai, December 1–2 · 15,000+ attendees from 130+ countriesGet tickets →
LIVE
BTC—ETH—SOL—BNB—XRP—ADA—AVAX—DOGE—LINK—DOT—MATIC—ATOM—LTC—TRX—TON—BTC—ETH—SOL—BNB—XRP—ADA—AVAX—DOGE—LINK—DOT—MATIC—ATOM—LTC—TRX—TON—
—▲0.0%
Web3

Agentic Payments Explained: When Software Pays Software

3 min readGuide
Agentic Payments Explained: When Software Pays Software

Agentic payments are payments made by an AI agent acting for you: it finds the thing, agrees the price and pays, and you see the result rather than the checkout. Instead of you clicking "buy", a program holds a limited permission you granted and spends inside it. Some of these systems run on ordinary card rails, and some settle in stablecoins onchain.

What stops an agent spending whatever it likes?

A permission slip, signed by you and checkable by the seller. Google's Agent Payments Protocol calls them mandates: you sign one saying what the agent may buy and on what conditions, and a second covering the exact payment once the basket is final. Mastercard and Visa have their own versions, where an agent is registered in advance and the card network refuses anything outside the limits attached to it. In the crypto version the limit lives in the wallet - the agent gets a budget and a key that can only spend that budget. It is the difference between handing someone your wallet and handing them a gift card with a set balance.

Who is building this?

Much of the payments industry at once, which is why the names overlap. Google's AP2, announced in September 2025 with more than sixty partners, handles the proof that you agreed. The Agentic Commerce Protocol from OpenAI and Stripe handles checkout, so an agent shopping inside a chat can buy from a merchant without pretending to be a person filling in a web form. Coinbase's x402 handles the crypto-native case, where a program pays a server directly in stablecoins. They behave as layers rather than rivals, and one purchase can use more than one.

Should you let one loose with your money?

Not with much of it yet. The standards are young, the rules about who pays for a mistake are still being written, and an agent that misreads a page will buy the wrong thing with complete confidence. The sensible version today is small and bounded: a cap you set yourself, a wallet or card holding only what you can afford to lose, and a record you can read afterwards.

If you want to try it, start with something cheap and repetitive - a paid data feed, not a flight - and keep the agent's money in its own wallet so a mistake has a ceiling. Next: what x402 is and machine-to-machine payment protocols.

← All guides