Agent Commerce Standards
The detailed standards, settlement, and regulation map for agent commerce.
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 |
Google: Agentic Commerce AI Tools & Protocol
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
When a human taps a card, the tap IS the consent. When an agent transacts, there is no tap. Verifiable Intent fills that gap — a cryptographic delegation chain binding agent actions to human-approved scope. Open-sourced by Mastercard and Google. See the full specification for the three-layer chain, eight constraint types, and verification protocol.
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
VI handles who authorized what. x402 handles how value moves. Together they sketch the stack for autonomous transactions where no human taps a card, but every transaction has a provable chain of consent.
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
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 |
| Peaq | Polkadot parachain | ~6s | Low | DePIN device-to-device, IoT micropayments |
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.
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 |
Failure Modes
Where an agent commerce standard breaks:
- Relocated extraction — a standard markets openness but a single patron still controls the routing table and clips every ticket.
- Consent gap — an agent transacts without a verifiable delegation chain, so a disputed charge has no provable authority behind it.
- Settlement mismatch — a card-tempo rail is bolted onto machine-tempo demand; latency and fees make agent-to-agent trade uneconomic.
- Regulatory stall — "human oversight" rules that assume a human in the loop make compliant real-time agent settlement impossible.
Context
- Agent Commerce — the hub this map supports
- Verifiable Intent — the consent layer in full
- Sui — settlement layer for machine-tempo trade
- Financialization — same extraction on different rails?
- Standards — what survives from experiment to adoption
Questions
When agents handle money, does the extraction move or disappear?
- If ACP and AP2 both succeed, does interoperability emerge or do agents pick sides?
- What breaks when regulation requires "human oversight" for transactions at machine tempo?
- Which patron extracts the least — and is that a stable equilibrium or a temporary truce?
Changes my mind: evidence that a single-patron protocol captures agent commerce while staying genuinely open and non-extractive would break the extraction-test thesis.
Next question: which layer — identity, execution, or settlement — locks in first, and does that decide who controls the routing table?