Skip to main content

Agents Can Buy Things Now: Inside the Visa + x402 + VGS Autonomous Commerce Stack

· 12 min read
Dora Noda
Dora Noda
Software Engineer

On April 8, 2026, an AI agent in San Francisco discovered a digital product through an API, evaluated three competing quotes, authorized a card payment, and took delivery of the asset — without a human ever touching a keyboard. That was the demo. The bigger story is the plumbing: Nevermined, Visa, Coinbase, and Very Good Security quietly stitched four separate stacks together into the first production system where an autonomous agent can move from discovery to settlement with zero human-in-the-loop checkpoints.

For two years, "agent commerce" has been a story of half-loops. PayPal's agent checkout still required a human tap to confirm. ERC-8183 kept agents trapped in crypto-native services. Visa Intelligent Commerce talked about card rails for agents but lacked a programmable settlement leg. Nevermined's announcement is the first time a single integration closes the loop — and it does so by bridging Visa's roughly 130 million merchant endpoints with HTTP-native stablecoin rails through a four-layer architecture that nobody, until now, had bothered to fuse.

The Four Layers, Demystified​

The stack is unglamorous in the way that load-bearing infrastructure usually is. Each layer does one thing, and the cleverness is in the joins.

Layer 1 — VGS as the data vault. Very Good Security operates the secure environment that captures, tokenizes, and routes cardholder data. The agent operator never touches a raw PAN. VGS's recent agentic commerce toolkit includes "Agentic API Pass-Through" — letting agents transact with PANs or network tokens without humans in the loop while preserving PCI compliance. By the time a payment leaves the vault, the actual card number has been swapped for a token that's useless to anyone who intercepts it.

Layer 2 — Visa Intelligent Commerce as the credential issuer. Visa Intelligent Commerce Connect, also launched April 8, is a network-, protocol-, and token-vault-agnostic on-ramp that generates secure payment credentials for an agent operating on a delegated user's behalf. Pilot partners include Aldar, AWS, Diddo, Highnote, Mesh, Payabli, and Sumvin. Crucially, Connect speaks four protocols in parallel — Visa's own Trusted Agent Protocol (TAP), Stripe's Machine Payments Protocol (MPP), Agentic Commerce Protocol (ACP), and Universal Commerce Protocol (UCP) — so merchants integrate once and accept agent-initiated payments across rails.

Layer 3 — Coinbase x402 as the machine-native point of sale. x402 revives the long-dormant HTTP 402 "Payment Required" status code. When an agent hits a paywall, the server replies with a 402, the agent attaches a signed stablecoin payment, and the request is retried — no API keys, no subscriptions, no checkout pages built for human eyes. As of March 2026, x402 had processed over 119 million transactions on Base and 35 million on Solana, handles roughly $600 million in annualized volume, and charges zero protocol fees. Cloudflare co-founded the x402 Foundation alongside Coinbase in September 2025; Stripe began routing USDC payments through it in February 2026.

Layer 4 — Nevermined as the orchestration brain. Nevermined sits on top, handling the parts the other layers don't: economic policy (budget caps, spend velocity), metering, settlement reconciliation, and the agent-identity primitives (Nevermined ID, signed DIDs) that prove which principal authorized which spend. Sign-up is reportedly five minutes. Sub-cent micropayments start at $0.001 — an order-of-magnitude cheaper than what traditional card processors can economically support per transaction.

Each piece existed before April 8. What's new is that they're now wired into a single integration path, with Visa shouldering the orchestration and PCI burden through its Acceptance Platform.

Why This Loop Was So Hard to Close​

To understand why the announcement matters, look at what failed before it.

PayPal's agent checkout, launched in 2025, gave agents a polished payment UX — but every transaction still surfaced to a human for tap-to-confirm. That's not autonomy; that's a fancier "Pay Now" button. ERC-8183 and other crypto-native agent commerce specs went the other direction: full autonomy, but only inside the wallet of services that accepted on-chain payments. The merchant universe accessible was a rounding error compared to Visa's network.

The actual production gap was a credential one. To pay a non-crypto merchant, an agent needs card credentials it can present at checkout. To pay a crypto-native API, an agent needs a stablecoin signing key. To do both safely on a user's behalf, it needs a vault that no one — including the agent operator running the LLM — can compromise. Visa Intelligent Commerce solved the first. x402 solved the second. VGS solved the third. Nobody had stitched them together with policy controls that a CFO would sign off on.

That's what Nevermined just did. The system gives agents persistent delegated spending authority — the agent doesn't ask permission for every $0.30 API call, but it operates inside a fence the user defined upfront. Guardrails include total budget limits, per-purchase caps, merchant allowlists, and time-based validity windows that auto-expire delegations. If the agent goes off-leash, the budget breaker trips before the damage gets meaningful.

The Guardrails Are the Product​

The technical achievement is impressive. The product achievement is the guardrails.

Every "should AI agents be allowed to spend money" debate eventually arrives at the same risk: an autonomous system with a credit card and a faulty objective function will, given enough time, find a way to spend more than it should. The fear is rational; the question has been who wears the loss when it happens.

Nevermined's stack pushes that risk into infrastructure rather than leaving it to the LLM's good judgment. Budget caps live in policy enforcement, not in a system prompt the agent could be talked out of. Per-purchase caps live in Visa's authorization layer, not in a check the agent does on itself. Merchant restrictions are enforced at credential issuance, not after the fact. Time-based validity is automatic — a delegation issued for "next 24 hours, $200 cap, AWS and Anthropic only" cannot be turned into "next month, $200,000 cap, anywhere" by clever prompt manipulation.

This is the same architectural lesson the early days of OAuth taught: trust the token, not the application. Nevermined's contribution is applying that principle to the agentic spending case, with cryptographic attestation that traces every charge back to the principal who authorized the policy in the first place.

The Protocol War, From a Distance​

Step back, and you can see Nevermined's integration is also a wager on which agent-payment protocols win.

Within 90 days of early 2026, the four major payment platforms each shipped their own standard. Visa unveiled the Trusted Agent Protocol (TAP). Google rallied 60 partners behind Agentic Payment Protocol (AP2). Coinbase shipped x402 with Cloudflare and Stripe support. PayPal announced Agent Ready. Stripe joined the parade on March 18, 2026, with the Machine Payments Protocol (MPP), co-authored with Tempo and aimed at microtransactions, pay-as-you-go, and recurring payments.

Each protocol has its own constituency. TAP optimizes for Visa's risk and identity primitives. AP2 leans on Google's identity rails. x402 is the crypto-native bet — HTTP-layer payments with stablecoin settlement and zero protocol fees. MPP courts SaaS pricing models on stablecoin rails. PayPal Agent Ready bets on its consumer trust and buyer protection moat.

Most builders see this as a war. Nevermined treats it as a tooling problem. Its protocol-first architecture supports A2A, x402, MCP, and AP2 natively, and the Visa Intelligent Commerce integration plugs into TAP, MPP, ACP, and UCP through Connect. The strategic position: don't pick a winner, become the orchestration layer that translates between them. If x402 wins, Nevermined wins. If TAP wins, Nevermined wins. If five protocols coexist and merchants need an aggregator, Nevermined wins twice.

The risk in that position is the same risk every aggregator faces — incumbents can vertically integrate and squeeze the middle. Stripe is building agent-payment functionality natively into its API. Coinbase is shipping x402 facilitators directly. Visa's Connect is itself an aggregation layer. Nevermined's defense is depth: tamper-proof metering, sub-cent micropayments, and identity primitives that the platforms haven't prioritized because they don't fit their existing risk models.

The $1.5 Trillion Question​

The market case is large enough to attract every aggregator with a pitch deck.

Juniper Research projects agentic commerce transaction volume will grow from $8 billion in 2026 to $1.5 trillion by 2030. Grand View pegs the market at $5.71 billion in 2025, $7.71 billion in 2026, and $65.47 billion by 2033. McKinsey's high-end forecast puts the global agentic commerce opportunity at $3 trillion to $5 trillion by 2030. VGS estimates AI agents could represent up to $385 billion of U.S. ecommerce spending by 2030.

Adoption is already non-trivial. The IBM Institute for Business Value reported in January 2026 that 45% of consumers use AI for at least part of their buying journey. Twenty-three percent of Americans made an AI-assisted purchase in the prior month. One in six 2025 Black Friday purchases were AI-assisted, and one in five Cyber Week orders involved an agent of some kind. Gartner expects that by the end of 2026, 25% of enterprise software purchases will involve AI agent mediation.

The contrarian view, and it's not crazy, comes from Stripe. Patrick and John Collison have written that agentic commerce is overhyped in the near term — that genuinely autonomous shopping will expand gradually rather than in a hockey stick. Stripe's current product positioning treats agents as augmenting human checkout, not replacing it. If Collison is right, Nevermined's stack is two years early. If he's wrong, it's right on time.

The honest read is somewhere in the middle. Pure consumer "buy this for me" autonomy will probably stay novelty until trust builds. But B2B and machine-to-machine flows — paying for API calls, compute, data, model inference, document generation — are already crossing into autonomous territory because the dollar amounts per transaction are small enough that the guardrail math works without much friction. That's the wedge Nevermined and x402 are exploiting: the $0.001-to-$5 transaction band where human approval is more expensive than the purchase.

What Comes Next​

Three things to watch over the next two quarters.

First, merchant adoption. Visa's Connect launch with seven pilot partners is the seed; the question is how fast the broader Visa Acceptance Platform integrates Connect. Until merchants accept agent-initiated payments natively, the long tail of agent purchases will route through workarounds. If the merchant ramp is fast, Nevermined's stack rides Visa's distribution. If it stalls, agents will continue to favor x402-native services where the loop is already closed.

Second, fraud and chargeback economics. Agent-initiated transactions create novel liability questions. Who's responsible when an agent gets prompt-injected into buying something the user didn't want? Visa, the merchant, the agent operator, the user? VGS's partnership with Forter on agentic-commerce fraud detection is one early answer. The legal answer will lag the technical one by 12-24 months.

Third, verticalization. Stripe, Coinbase, and Visa are all building competing aggregation layers. If any of them ship orchestration parity with what Nevermined sells today, the middle gets squeezed. Nevermined's bet is that the metering, identity, and policy primitives are deep enough to be a defensible niche even if the rails commoditize. That bet either pays off in 18 months or it doesn't.

The April 8 announcement is the kind of integration that won't trend on Twitter for long. But it's the one that future "agentic commerce" case studies will reference as the day the loop closed. Discovery, evaluation, payment, delivery — all programmatic, all enforceable, all auditable. The agent economy needed a payment stack it could grow into. As of this month, it has one.

BlockEden.xyz provides enterprise-grade RPC, indexing, and identity infrastructure for the agent economy across Sui, Aptos, Ethereum, and ten other chains. As autonomous agents become first-class consumers of on-chain APIs, explore our API marketplace to build on the rails that agent commerce will run on.

Sources​

Share on Twitter