AI Agents Ship Payment Protocols While Governance Stalls

On April 2, Coinbase transferred the x402 protocol to the Linux Foundation, backed by a founding membership that includes AWS, Google, Microsoft, Visa, Mastercard, American Express, Shopify, and Circle. Two weeks before that, Stripe and blockchain startup Tempo launched the Machine Payments Protocol, an open standard for AI agent payments in both fiat and cryptocurrency — with Tempo building the underlying blockchain infrastructure. And Google's Agent Payments Protocol, AP2, introduced in late 2025, was already being adopted by dozens of organizations including Mastercard, PayPal, and Adyen.

The companies behind these protocols — Visa, Mastercard, Stripe, Google, AWS, Microsoft — process the majority of global card transactions and host the majority of cloud computing workloads. When they converge on agentic payments as a priority, through three competing approaches, autonomous agent commerce has moved from speculation to active capital deployment.

Yet the AI governance frameworks meant to regulate this buildout are running behind. NIST launched an AI Agent Standards Initiative in February, but substantive deliverables aren't expected before late this year. As documented in a recent wave of 19 AI laws in just two weeks, legislation is multiplying at the state level without coordination. The EU AI Act reaches full enforcement for high-risk systems in August. The enterprises deploying agents are making protocol integration decisions now, without a shared rulebook.

Three Architectures, One Market

The three protocols aren't replacements for each other. They address different parts of what's likely to become a single agentic commerce system.

x402 works at the HTTP layer. It repurposes the long-dormant 402 status code — "Payment Required" — to embed stablecoin settlement directly into web requests. An agent hits a paid endpoint, receives a 402 response containing payment terms, and settles in USDC on Base or Solana. No checkout page, no session, no human present. Coinbase built it; the Linux Foundation now governs its development as a vendor-neutral standard.

Stripe's Machine Payments Protocol takes a different approach. MPP supports hybrid payment methods — fiat and crypto — through what Stripe calls Shared Payment Tokens, letting agents pay with cards, buy-now-pay-later, or stablecoins depending on the merchant. Where x402 is crypto-native and settles on-chain, MPP bridges existing financial AI infrastructure — making machine payments accessible through familiar rails. Stripe users accept MPP payments through the same API they already use, and services ranging from headless browser sessions to physical mail delivery were available in MPP's directory at launch.

Visa is supporting MPP as a design partner, developing specifications for card-based AI agent payments and extending the protocol to work on the Visa Acceptance Platform. "We're entering a moment where agents can make decisions, move resources and pay for services on their own," said Rubail Birwadker, Visa's global head of growth products. "But for these kinds of payments to scale, security has to be built into every layer."

AP2 addresses a different problem entirely. Google's protocol doesn't move money — it governs authorization. Using Verifiable Credentials called "Mandates," AP2 creates cryptographic proof that a human authorized an agent to spend within specific parameters. It supports two modes: Cart Mandates for real-time transactions where the human is present, and Intent Mandates for delegated tasks where the agent operates independently. The distinction matters: x402 handles settlement, AP2 proves someone approved the spending, and MPP bridges payment methods. A single transaction could theoretically involve all three.

That theoretical interoperability hasn't been built. Alchemy acknowledged as much when it launched AgentPay on April 8, positioning the product as an AI interoperability layer between protocols. "It's only going to get more fragmented as more systems launch," said Guillaume Poncin, Alchemy's CTO. That an interoperability product was necessary within days of the protocols launching says something about where the ecosystem stands. This fragmentation echoes an earlier pattern, when three agent payment systems launched without interoperability.

The First Regulated Agent Payment

While protocols were being announced, a regulated transaction was being completed. In March, Santander and Mastercard executed what they described as Europe's first live end-to-end payment by an AI agent, using Mastercard's Agent Pay program within Santander's banking infrastructure.

The mechanics worked. An AI agent authorized a payment, executed it through a regulated bank's live systems within predefined limits and customer-granted permissions, and the transaction settled. Matias Sanchez, Santander's global head of cards and digital solutions, said the bank's role is to shape innovation responsibly — embedding security, governance, and customer protection into the design from the start.

What the pilot doesn't resolve — and what none of the three protocols fully address — is accountability. Who bears liability when an agent's payment goes wrong? Which regulatory framework applies when autonomous agents operate across jurisdictions? How are disputes resolved when no human was present for the transaction?

Mastercard has been building toward answers. In early March, it introduced Verifiable Intent, an open-source framework that creates a cryptographic audit trail linking consumer identity, instructions, and transaction outcomes into a tamper-resistant record. "As autonomy increases, trust cannot be implied. It must be proven," said Pablo Fourez, Mastercard's Chief Digital Officer. The framework is designed to interoperate with AP2 and could provide the accountability mechanism that payment protocols themselves lack.

The Governance Deficit

This AI infrastructure is being deployed into enterprises that adopted AI agents far faster than they built governance around them.

OutSystems surveyed 1,900 IT leaders globally for its 2026 State of AI Development report and found that 96% of organizations are already using AI agents in some capacity. The adoption curve didn't gradually steepen — it collapsed. AI agents went from pilot projects to operational AI tools before most companies had policies, oversight structures, or centralized management in place.

The consequences of that speed show in how agents are actually managed. Most organizations rely on ad hoc oversight — team-level controls and inherited IT policies designed for a different era. Only 12% have implemented a centralized platform to manage their agent deployments, despite near-universal concern about the complexity and security risk agents introduce.

NIST recognized this deficit when it launched the AI Agent Standards Initiative in February. The initiative identifies specific agent security threats — prompt injection, excessive write access, untrusted resource interaction — and recommends enterprise-grade agent identities, least-privilege access, and complete audit trails.

Research from the Cloud Security Alliance draws a sharper line. In a recent survey of security leaders, the vast majority reported lacking full visibility into their AI agent identities — a finding consistent with broader research showing nearly half of organizations cannot see their AI agent traffic. Most said they doubted their organization's ability to detect or contain a compromised agent. The gap between agent deployment and agent security is widest at identity and access management — the exact area where agentic payment protocols create the most new exposure.

Some of NIST's recommendations are already running in production. AgentPMT, an AI agent platform that operates a marketplace and autonomous payment system for AI agents, has built per-agent budget controls with smart-contract-enforced spend limits, full audit trails for every agent interaction, and human-in-the-loop approval via biometric mobile authentication. Its x402Direct system settles payments autonomously while the smart contract blocks any spend outside predefined parameters — a working implementation of the kind of AI governance controls NIST describes.

Building for Protocol Uncertainty

The three protocols may converge into complementary layers: x402 for settlement, AP2 for authorization, MPP for payment method bridging. Alchemy is already building those connections. But convergence, if it happens, will take years. For a condensed view of this landscape, see our overview of how AI agent payment protocols are outpacing governance standards.

Companies making AI infrastructure decisions today face a practical choice about where to invest. The protocols will evolve, merge, or fork. The governance requirements won't change. Budget controls, agent identity management, audit trails, and human oversight are foundational regardless of which payment protocol prevails.

AgentPMT's architecture reflects this approach: build on x402 for settlement through x402Direct's smart contracts, but wrap it in governance controls — spend caps, authorized tool categories, per-agent budgets, and real-time logging — that work independently of which protocol handles the transaction. The settlement mechanism can change; the controls have to stay.

For engineering teams, this means building payment integrations that don't lock into a single protocol. For finance teams, it means establishing spend controls before autonomous agents start transacting independently. For security teams, it means treating agent identities with the same rigor applied to human users — access management, credential rotation, continuous monitoring.

The payment infrastructure for AI agents is being assembled in real time, by competing teams building on different technical assumptions toward the same commercial reality — a maturing agentic economy where machines transact on behalf of humans. Budget controls, identity standards, audit mechanisms, and human oversight can't wait for the protocols to sort themselves out — and the companies building them aren't waiting.

---

Sources

• Linux Foundation x402 Foundation Launch — Linux Foundation
• Introducing the Machine Payments Protocol — Stripe
• Stripe-Backed Crypto Startup Tempo Releases AI Payments Protocol — Fortune
• Google's AP2: A New Protocol for AI Agent Payments — Vellum AI
• Alchemy Debuts Tool to Provide Interoperability for Agentic Payments — PYMNTS
• Mastercard and Santander Complete Europe's First AI Agent Payment — PYMNTS
• Announcing the AI Agent Standards Initiative — NIST
• Agentic AI Goes Mainstream in the Enterprise — OutSystems
• Digital Money Has a New Payment Standard — PYMNTS
• Mastercard Unveils Open Standard to Verify AI Agent Transactions — PYMNTS
• x402 vs Stripe MPP: How to Choose Payment Infrastructure for AI Agents — WorkOS
• Visa Scales Agentic Commerce Through Stripe Protocol Collaboration — PYMNTS
• Why Stripe's Machine Payments Protocol Signals a Turning Point for Micropayments — Forrester
• The AI Agent Governance Framework Gap — Cloud Security Alliance