Agent Commerce
What happens when AI agents handle payment and banks become optional?
Agent commerce is the emerging standard for AI agents transacting autonomously. The agentic economy is projected to reach $3-5 trillion by 2030 — and the standards being set now determine who captures that value.
The Standards War
Three protocol families are competing to become the default rails for agent transactions:
| Protocol | Backed By | Focus | Status |
|---|---|---|---|
| ACP (Agentic Commerce Protocol) | OpenAI + Stripe | Agent-to-merchant commerce | PayPal adopting for 2026 |
| AP2 (Agent Payments Protocol) | Agent-led payments | 60+ partners incl. Mastercard, AmEx, Coinbase, PayPal. Sui as launch partner | |
| x402 | Coinbase + Google + Ethereum Foundation | Crypto-native agent payments | 50M+ transactions, Stripe on Base. Mastercard Verifiable Intent complements as auth layer |
Plus the communication layer underneath:
| Protocol | Function | Analogy |
|---|---|---|
| A2A (Google) | Agent-to-agent coordination | TCP — establishing connections |
| MCP (Anthropic → Linux Foundation) | Agent-to-tool access | HTTP — accessing resources |
| AP2 / ACP / x402 | Agent-to-payment | SMTP — delivering the message |
Verifiable Intent
Mastercard and Google co-developed Verifiable Intent — a cryptographic, tamper-resistant record linking consumer identity, their instructions, and the resulting transaction. It solves the problem AI agents create: no visible moment of human confirmation like tapping a card.
| Property | What It Does |
|---|---|
| Cryptographic proof | Links consumer identity → instructions → transaction in a tamper-proof chain |
| Selective disclosure | Only minimum necessary data shared, only when needed |
| Multi-party verifiable | Consumer, merchant, and issuer can all verify what was authorized |
| Open standards | Built on FIDO, EMVCo, IETF, W3C — protocol-agnostic by design |
| Dispute resolution | Clear audit trail replaces guesswork in chargebacks |
The emerging agentic payments stack:
IDENTITY/AUTH Verifiable Intent (Mastercard + Google)
↓
PAYMENT EXECUTION x402 (Coinbase + Google + Ethereum Foundation)
↓
SETTLEMENT Stablecoin rails (Sui, Base, Solana)
↓
FEEDBACK Reputation, on-chain attestations
Verifiable Intent handles who authorized what. x402 handles how value moves. Together they sketch the stack for trillions of micropayments across billions of agents — where no human taps a card, but every transaction has a provable chain of consent.
Partners: Adyen, Fiserv, Worldpay, IBM, Checkout.com, Basis Theory. Spec open-sourced at verifiableintent-dev. Integration into Mastercard Agent Pay APIs coming 2026.
Extraction test score: Green flags — open-sourced spec, built on existing open standards (FIDO/W3C/EMVCo), privacy-preserving by design. Watch: Mastercard still controls the Agent Pay API layer.
Bank Bypass
20% of card-based settlement could be displaced by AI agents and stablecoins by end of 2026. The mechanism:
TODAY: Human → App → Bank → Network → Bank → Merchant
AGENT: Agent → Stablecoin → Settlement → Done
| What Banks Provide | Agent Alternative |
|---|---|
| Trust (identity verification) | Verifiable Intent, cryptographic proof, Visa Trusted Agent Protocol |
| Routing (payment networks) | x402, stablecoin rails, solver networks |
| Compliance (KYC/AML) | Modular compliance layers, on-chain identity |
| Settlement (clearing) | Atomic settlement, smart contracts |
| Credit (lending) | DeFi lending, tokenized credit |
85% of financial institutions believe current systems are insufficient for high-volume agent transactions. Visa and Mastercard responded by embedding stablecoin rails into their networks — adapting, not dying.
The question isn't whether agents bypass banks. It's whether the new intermediaries are better or just different rails running the same extraction.
The Extraction Test
Every agent commerce standard must answer: does this standard reduce extraction or relocate it?
| Standard | Extraction Risk | Mitigation |
|---|---|---|
| ACP (Stripe) | Merchants remain merchant of record — Stripe still clips the ticket | Open standard, competing PSPs |
| AP2 (Google) | Google controls the protocol definition | Apache 2.0 license, 60+ partners |
| x402 (Coinbase) | Coinbase as primary implementation | Chain-agnostic design, open source |
| Virtuals ACP | Protocol token creates insider dynamics | 18,000+ agents, $470M+ GDP — real volume |
The honest answer: every standard has a patron. The financialization pattern doesn't disappear because agents handle the money. It disappears when the standard is open enough that no single gatekeeper controls the routing table.
Apply to any agent commerce standard:
| Question | Red Flag | Green Flag |
|---|---|---|
| Who controls the routing table? | Single company | Open, forkable protocol |
| Can you switch providers without rebuilding? | Lock-in by design | Interoperable by default |
| Is the source auditable? | Closed implementation | Open source, on-chain |
| Where does the fee go? | Platform extraction | Network participants |
The Four Phases
Regardless of which standard wins, agent transactions follow the same loop:
REQUEST → NEGOTIATION → TRANSACTION → EVALUATION
↑ ↓
└──────────── feedback loop ───────────┘
| Phase | What Happens | Who Does It |
|---|---|---|
| Request | Agent formulates need | AI intent system |
| Negotiation | Terms agreed, proof created | Agent-to-agent (A2A) or agent-to-merchant (ACP) |
| Transaction | Value transfers | Stablecoin rails (x402), card rails (AP2), or smart contracts |
| Evaluation | Quality assessed, reputation updated | Evaluator agents, on-chain attestations |
The evaluation phase is what makes this a VVFL — each transaction teaches the next one. Agents that transact well get more business. Agents that extract get excluded.
The Numbers
Industry projections — treat as directional, not precise.
| Metric | Value | Source | Confidence |
|---|---|---|---|
| Agentic economy by 2030 | $3-5 trillion | AWS | Projection |
| Growth rate | 45% CAGR | IBM | Projection |
| Agent-driven online spending | 50%+ ($1T+) | Nevermined | Projection |
| x402 transactions processed | 50M+ | Chainstack | On-chain |
| Virtuals agent network | 18,000+ agents | Virtuals | On-chain |
| Card displacement by 2026 | ~20% | PaymentsSource | Opinion |
Settlement Infrastructure
The payment protocols above describe negotiation. Settlement is where value actually moves.
| Protocol | Settlement Layer | Speed | Cost | Best For |
|---|---|---|---|---|
| ACP | Card rails (Stripe) | 1-3 days | 1-2% | Merchant checkout |
| AP2 | Card or stablecoin | Variable | 0.5-2% | Multi-rail flexibility |
| x402 | Stablecoin on-chain | ~390ms (Sui) / ~400ms (Solana) | <0.1% | Agent-to-agent, DePIN |
Why the settlement layer matters: When an AI agent buys something, it needs to settle instantly, cheaply, across borders. Traditional rails can't do that at machine tempo. Stablecoins on high-throughput chains close the gap between intent and finality.
The Libra lineage: Sui's founders built Move at Meta's Diem project. Google chose Sui as AP2 launch partner. Meta is now sending stablecoin RFPs with Stripe as likely pilot. The team that built Meta's payment infrastructure rebuilt it permissionless — and Meta is buying a seat on the rails their own engineers created.
The Three Flows pattern applies: INTENT → ROUTE → INFRASTRUCTURE → SETTLE → FEEDBACK. Agent commerce is the intent and routing layer. Stablecoins on Sui/Solana are the settlement layer. The VVFL feedback loop teaches each transaction to improve the next.
Regulation
| Jurisdiction | Status | Key Rule |
|---|---|---|
| EU | AI Act high-risk deadline Aug 2026 | Traceability, explainability, human oversight for financial AI |
| US | CFPB: no tech exemptions | Consumer protection laws apply to AI agents |
| Direction | Continuous auditing by 2027 | Real-time monitoring replaces model transparency |
Context
- Three Flows — Messages, money, data: same architecture, same settlement pattern
- Sui — Settlement layer: ~390ms finality, PTB batching, Move safety
- Stablecoins — The settlement instrument
- Payments Industry — $25T B2B cross-border friction meets stablecoin efficiency
- Phygital Beings — The actors who transact: devices, agents, and humans wearing different hats
- A2A Protocol — How agents coordinate before they transact
- UCP — Google's Universal Commerce Protocol — the commerce operations layer
- AP2 — Google's Agent Payments Protocol — payment authorization via VDCs
- Payment Rails — Intent-based payment infrastructure
- Payments Industry — $25T B2B cross-border friction meets agentic settlement
- A&ID Template — Visual language for agent-instrument systems
- Financialization — Same game on different rails?
- Standards — What survives from experiment to adoption
- Protocols — Rules for making progress
Questions
When agents handle money, does the extraction move or disappear?
- If ACP (OpenAI + Stripe) and AP2 (Google) both succeed, does interoperability emerge or do agents pick sides?
- What breaks when regulation requires "human oversight" for transactions that happen at machine tempo?
- Is the honest answer that every standard has a patron — and the question is which patron extracts the least?