Agentic commerce is the model in which autonomous AI agents act as proxies for buyers, executing the full purchase lifecycle (discovery, authorization, payment, fulfillment) from a goal rather than a click. Instead of a person browsing a site and pressing "buy," an agent reads an instruction like "order trail-running shoes under $150 that arrive Friday," evaluates options across merchants, and completes the transaction inside a wallet, a chat surface, or a checkout protocol. The model is moving fast: Adobe Analytics measured a 4,700% year-over-year jump in generative-AI traffic to US retail sites between July 2024 and July 2025, and the Coinbase-led x402 payment protocol processed roughly 165 million agent transactions in its first months. This guide covers the working definition, the four-step mechanism, the seven-layer infrastructure stack, the six commerce protocols competing to power it, the live deployments shipping today, and why stablecoins have become the default settlement currency for machine-to-machine payments.
​

Agentic commerce is a category of online purchasing where an AI agent acts as the buyer, holding delegated authority to discover products, authorize spending within set limits, and execute payments on a user's behalf. McKinsey projects $3 to $5 trillion in agent-orchestrated retail spend by 2030, with consumer, enterprise, and machine-to-machine variants all in production.
​

Agentic commerce is a category of e-commerce where the buyer is an AI agent operating with delegated authority. The agent receives a goal in natural language, plans the steps required to satisfy it, queries merchants or marketplaces, evaluates options against the user's constraints, authorizes a payment within preset limits, and triggers fulfillment. Unlike a recommendation engine or a chat assistant, an agent transacts: it moves money and produces an order at the end of the conversation.
​

The category has three working forms in 2026. Consumer-side agents (OpenAI's ChatGPT with Instant Checkout, Perplexity's Comet browser, Amazon's Rufus, Google AI Mode shopping) initiate purchases on behalf of an individual user. Enterprise-side agents (Microsoft Copilot, Salesforce Agentforce, custom internal agents) reorder office supplies, renew SaaS subscriptions, and pay invoices on behalf of a company. Machine-to-machine agents pay micro-amounts for API calls, compute, and data feeds using protocols designed for sub-cent transactions.
​

The market sizing converges around a number large enough to matter. McKinsey's QuantumBlack October 2025 study projects that agentic commerce will orchestrate $3 trillion to $5 trillion in global retail spend by 2030, with up to $1 trillion of that in US business-to-consumer alone. Juniper Research's April 2026 forecast takes a tighter view, projecting $8 billion in agentic spend in 2026 climbing to $1.5 trillion globally by 2030. Both numbers describe a five-to-six-year compounding curve, not a near-term inflection.
​

Agentic commerce works through a four-step lifecycle: discovery, authorization, payment, and fulfillment. The agent locates merchants and products, proves the user delegated the purchase, moves funds through a card rail or a stablecoin protocol, and triggers the merchant's existing fulfillment pipeline. Each step uses a distinct protocol, and the protocols are designed to compose end-to-end.
​

Every agent purchase moves through four steps. Discovery comes first, then authorization, then payment, then fulfillment. Each step has a different set of protocols and a different set of failure modes.
​

Discovery. The agent learns what merchants exist, what products they sell, and what payment methods they accept. Two protocols dominate this layer. Anthropic's Model Context Protocol (MCP), open-sourced in November 2024, lets an agent query any compliant server for tools and data over JSON-RPC 2.0. Google's Universal Commerce Protocol (UCP), announced at NRF 2026, exposes merchant capability profiles so an agent can list supported services (checkout, product search, returns) before it transacts.
​

Authorization. The agent proves to the merchant that a real user delegated the purchase, and the merchant proves to the network that the agent is legitimate. This is the layer where most agent-fraud risk concentrates, and it is the most contested protocol layer. Google's Agent Payments Protocol (AP2) introduces cryptographic "mandates" — signed digital receipts where the user attests to a specific transaction intent. Visa's Trusted Agent Protocol (TAP), launched October 14, 2025 with Cloudflare, signs the agent's identity into HTTP request headers; merchants verify the signature against Visa's directory.
​

Payment. The agent moves funds. For card rails, the dominant pattern is a scoped ephemeral credential. OpenAI's Agentic Commerce Protocol (ACP), co-developed with Stripe, issues a Shared Payment Token (SPT) bound to a specific merchant and a specific dollar amount, time-bounded, single-use. Mastercard's Agentic Tokens (part of Agent Pay) operate similarly. For onchain rails, agents pay in stablecoins over Coinbase's x402 protocol or Stripe's Machine Payments Protocol (MPP).
​

Fulfillment. The merchant runs the order through its existing pipeline (warehouse, logistics, customer service). Agentic commerce does not replace fulfillment infrastructure; it sits in front of it. The merchant retains the relationship with the buyer for refunds and dispute handling, plus post-purchase support, even when an agent placed the order on the buyer's behalf.
​

The mechanism only works when each step composes with the next. A shopping agent that runs MCP for discovery still needs ACP or UCP for the checkout call, AP2 or TAP for the authorization signature, and a settlement rail (card network, Stripe, x402) for the actual money movement. The protocols are designed to compose, but the composition is still being negotiated in real time.
​

The agentic commerce stack is a seven-layer reference taxonomy spanning AI platforms, protocols, payments, card issuance, checkout execution, merchant enablement, and trust and security. Each layer maps to a distinct vendor cluster, from OpenAI and Google at the surface layer to Cloudflare and HUMAN Security at the trust layer, with well-funded startups across the middle four.
​

Rye, a checkout-execution startup, published a 7-layer framework for the agentic commerce stack that has become the most-cited reference taxonomy for the category. Each layer maps to a different set of incumbents and a different cluster of well-funded startups.
​

Where shopping journeys originate. The named platforms are OpenAI (ChatGPT), Google (AI Mode, Gemini), Anthropic (Claude), Perplexity (Comet browser), Amazon (Rufus), Microsoft (Copilot), and Meta. Each surface decides which merchants its agents can transact with and which protocols it supports.
​

The communication layer. ACP (OpenAI + Stripe), UCP (Google + Shopify), AP2 (Google + payment networks), MCP (Anthropic, donated to the Linux Foundation's Agentic AI Foundation in December 2025), A2A (Google), and Visa TAP all sit here. Adoption is the open question: most merchants will need to support more than one.
​

The financial rails. Network incumbents (Visa, Mastercard, PayPal) and processors (Stripe, Adyen, Checkout.com) compete with a well-funded startup tier. Basis Theory raised a $33 million Series B in October 2025 led by Costanoa Ventures. Skyfire has raised $9.5 million from Neuberger Berman, a16z CSX, and Coinbase Ventures. Nekuda raised a $5 million seed in May 2025 from Madrona, with Amex Ventures and Visa Ventures participating. The combined total across just these three is roughly $47.5 million, concentrated entirely on agent-payment infrastructure.
​

Programmable virtual cards that an agent can spin up per merchant or per session. Lithic, Stripe Issuing, Marqeta, Highnote, Ramp, and Brex are the named players. Card issuance is the bridge between agentic protocols and the existing card-network universe.
​

The "browser robot" layer. When a merchant has not adopted ACP or UCP, an agent still needs a way to complete the purchase. Rye, Induced AI, Henry Labs, and other vendors run automation that fills out checkout forms on the merchant's actual site. Rye itself raised $14 million.
​

The product-data and catalog layer. Shopify, commercetools, BigCommerce, Adobe, and SAP supply the e-commerce backends; specialists like Salsify, Syndigo, and Akeneo standardize the product feeds that agents read.
​

Fraud, bot detection, and agent verification. HUMAN Security, Riskified, Forter, Signifyd, Kasada, DataDome, Akamai, and Cloudflare sit here. The Agentic Commerce Consortium, formed by Basis Theory with Lithic, Crossmint, Skyfire, Rye, Channel3, and others, is publishing security standards across this layer.
​

Insignia Business Review's April 2026 analysis argues that durable value will accrue to the settlement rails (stablecoin issuers, payment facilitators) rather than the open-source protocols themselves, on the same logic that TCP/IP captured no rents while Akamai, Cloudflare, and AWS captured plenty. That thesis is contested but informs how strategic acquirers are positioning.
​

Six protocols define the working stack as of April 2026: ACP from OpenAI and Stripe, UCP from Google, AP2 from Google with payment networks, MCP from Anthropic, A2A from Google, and Visa TAP. Each handles a different slice of the discovery, authorization, payment, or trust layer, and most production deployments compose two or three together.
​

Six protocols define the agentic commerce stack as of April 2026. Most production deployments use two or three together. The table below summarizes backers, scope, and current status.

ACP is the protocol that powers ChatGPT Instant Checkout. The core primitive is the Shared Payment Token: instead of exposing the buyer's saved card to the agent, the buyer's payment provider issues a token scoped to a specific merchant and a specific amount, with a short expiry. The agent passes the token to the merchant; the merchant runs it through its normal processor (typically Stripe). OpenAI launched ACP on September 29, 2025 with Etsy as the first integrated merchant; Shopify integration was announced but onboarding stalled. Only about a dozen Shopify merchants went live before OpenAI retired Instant Checkout in March 2026 in favor of dedicated retailer apps (Walmart, Target, and Instacart) inside ChatGPT.
​

UCP is Google's answer to ACP, designed to span the full journey rather than the checkout call. Sundar Pichai announced UCP at NRF 2026 in his January keynote, with backing from Shopify, Walmart, Target, Best Buy, Etsy, PayPal, Visa, Stripe, and American Express. UCP composes with AP2 for payments and MCP for tool access, which is part of why Google positions it as a "universal" rather than a competitor to ACP. The technical scope covers product discovery and capability negotiation, plus checkout and post-purchase handoff.
​

AP2 separates the question of "did the user authorize this?" from the question of "is the payment valid?" Each authorization is a cryptographic mandate signed by the user (typically by the user's wallet or AP2-compatible identity provider). The agent presents the mandate at checkout. The mandate carries the merchant identifier plus the transaction amount and expiry. The merchant verifies the signature, runs the payment, and stores the mandate as a non-repudiable receipt. AP2 was developed with PayPal and other payment partners; it is the protocol UCP defaults to for the payment step.
​

MCP is the protocol that makes agents into shoppers in the first place. Anthropic open-sourced it in November 2024; it borrows transport ideas from the Language Server Protocol and runs over JSON-RPC 2.0. An agent uses MCP to discover and call tools (search a catalog, query an inventory database, read a return policy) exposed by any compliant server. Google, Microsoft, OpenAI, Visa, and Mastercard all committed to MCP within six months of release. Anthropic donated MCP to the Agentic AI Foundation at the Linux Foundation in December 2025, co-founded with Block and OpenAI. MCP is now upstream of every commerce protocol in active use.
​

A2A is Google's protocol for one agent delegating to another. A shopping agent might delegate price comparison to a specialist agent, or delegate the actual checkout to a merchant-side agent. By April 2026, A2A had reached production deployment at over 150 organizations across Microsoft Azure AI Foundry, Amazon Bedrock AgentCore, and Salesforce, per the Linux Foundation's April 2026 release.
​

TAP solves a problem that the merchant side cares about: distinguishing a legitimate agent from a bot or a scraper. TAP signs the agent's identity into HTTP request headers using public-key cryptography. The merchant fetches Visa's directory key and verifies the signature, confirming both the agent and the user authorization. Visa launched TAP on October 14, 2025 with Cloudflare; the specification is on GitHub. Early implementers include Adyen, Ant International, Checkout.com, Coinbase, Fiserv, Microsoft, Shopify, Stripe, and Worldpay.
​

Agentic commerce ships in five concrete production deployments as of April 2026: ChatGPT Instant Checkout, Amazon Buy for Me, Mastercard Agent Pay, Visa Intelligent Commerce, and Coinbase Agent.market with x402. Three settle on card rails, two on stablecoins, and one (Amazon Rufus) is on a $12 billion incremental annualized run-rate per Q4 2025 earnings.
​

Agentic commerce is past the demo stage in five concrete deployments worth naming.
​

ChatGPT Instant Checkout (OpenAI + Etsy + Stripe). The first mainstream consumer agent-checkout flow. A US ChatGPT user types "find me a handmade ceramic mug under $40," and the agent returns Etsy listings the user can buy without leaving the chat. Stripe processes the payment via the ACP shared payment token. Adoption was thin. Only about a dozen Shopify merchants ever went live alongside the original Etsy integration before OpenAI retired Instant Checkout in March 2026, replacing it with dedicated retailer apps inside ChatGPT (Walmart, Target, and Instacart were the launch partners).
​

Amazon Buy for Me (Amazon Nova + Anthropic Claude). When a buyer searches Amazon for a branded item Amazon does not stock, Rufus offers to buy it from the brand's site directly. Buy for Me went from 65,000 SKUs at its April 2025 launch to over 500,000 by late 2025. Amazon's Q4 2025 earnings disclosed that Rufus reached more than 300 million customers and was on a roughly $12 billion incremental annualized sales run-rate by year-end.
​

Mastercard Agent Pay. The first major card network to enable all of its US cardholders for agent-initiated payments. Citi and US Bank cardholders gained first access in September 2025; Mastercard announced full US rollout in November 2025, with global expansion through 2026. Agent Pay relies on Mastercard Agentic Tokens, scoped per agent and per session.
​

Visa Intelligent Commerce and the Agentic Ready Program. Visa's Intelligent Commerce initiative rolled the Agentic Ready Program out to banks across Europe and the UK first, then expanded to Asia Pacific and Latin America. Pilot partners include Skyfire, Nekuda, PayOS, and Ramp; the partnerships have completed end-to-end consumer and B2B agent purchases in closed beta. Intelligent Commerce Connect, announced separately, lets businesses bridge into agentic commerce without changing their existing processor stack.
​

Coinbase Agent.market and x402. Coinbase launched Agent.market in April 2026 as an app store for autonomous agents that pay each other in stablecoins via x402. By late April 2026, x402 had 69,000 active agents, 165 million transactions, and roughly $50 million in cumulative volume — an average value per call of about 30 cents, calibrated for sub-cent micropayments rather than retail purchases.
​

The shape of these deployments matters as much as the count. Three of the five run on traditional card rails; two run on stablecoins. All five run on top of one or more of the six protocols above. None of them have replaced human checkout at meaningful volume yet, but Juniper's $8 billion 2026 estimate for global agentic spend is consistent with the early-pilot footprint.
​

Agentic commerce promises broader merchant comparison, completed transactions, and machine-to-machine billing without API key signups. The risks split into four buckets: authorization (did the user approve?), identity (is this an agent or a scraper?), fraud (who owns the chargeback?), and merchant discovery (algorithmic ranking can collapse organic channels). Juniper's April 2026 study identified trust as the number-one barrier.
​

The benefits side of agentic commerce gets the most attention because the demos are vivid. The risks side is where most production work in 2026 is actually happening.
​

The benefits. An agent can compare prices across more merchants in seconds than a human will compare in an afternoon. Adobe's consumer survey of 5,000 US shoppers found 38% had already used generative AI for shopping by mid-2025, and 52% planned to in the following year, with the most-cited tasks being research (53%), product recommendations (40%), and deal-finding (36%). For merchants, an agent that completes a transaction at all is more valuable than a human visit that bounces; Adobe Analytics measured AI-driven revenue-per-visit climbing 84% between January and July 2025. For machine workloads, x402 lets a server bill an agent in dollars without an API key signup, a payment-method form, or a monthly invoice cycle.
​

The risks. Three categories of risk are getting worked on simultaneously. First, authorization risk: how does a merchant know the user actually approved this purchase? AP2 mandates and Visa TAP signatures are designed to give merchants a non-repudiable answer. Second, identity risk: how does a merchant tell an agent from a scraper or a bot? TAP's signed-headers approach is one answer; trust-and-security vendors at Layer 7 of the Rye stack handle the rest. Third, fraud and dispute risk: chargeback rules assume a human pressed the buy button. Mastercard's Agentic Tokens and ACP's Shared Payment Tokens both encode the agent identity into the transaction record so that disputes can attribute correctly. Juniper Research's April 2026 study explicitly identified trust as the number-one barrier to agentic commerce deployment, ahead of all technical concerns.
​

A fourth risk lives at the merchant rather than the network. If an agent surface ranks products algorithmically, the merchant's organic discovery channel can collapse overnight. Most large retailers have responded by integrating into ACP, UCP, and AP2 in parallel rather than picking one, on the same logic that drove early multi-channel strategies in mobile commerce.
​

AI agents need dedicated payment protocols because card rails fail at machine scale: interchange fees compress at sub-cent ticket sizes, chargebacks assume a human dispute window, and settlement runs on banking hours. Stablecoin protocols like x402 settle in seconds for cents in gas, regardless of ticket size, against the same dollar token across chains.
​

Most agent purchases on consumer surfaces still settle on card rails. Behind the surface, a different model is taking hold for machine-to-machine and B2B agent payments: stablecoin settlement over open protocols. The reasons are mechanical, not ideological.
​

Card rails were designed for human-scale transactions. Interchange and processor fees compress at low ticket sizes; chargebacks assume a human dispute window; settlement runs on banking hours. None of those constraints fit an agent that wants to pay fractions of a cent for an API call ten thousand times an hour. x402 was designed exactly for this regime: a server returns HTTP 402 with a payment-required header listing the chain identifier (CAIP-2), the token contract, the amount, and the recipient address; the agent signs an EIP-3009 or Permit2 payload and returns it via a payment-signature header; the server verifies through a facilitator and returns the resource. The whole flow runs over standard HTTP with no separate API key handshake.
​

The settlement currency is almost always a dollar-denominated stablecoin. Stripe added x402 support in February 2026 with USDC on Base as the launch pair, and it has signaled support for additional chains and tokens. Coinbase's x402 docs cover EVM chains plus Solana out of the box. The economics are straightforward: at the time of writing, USDC supply sits at roughly $77 billion and USDT at roughly $189 billion (per DeFiLlama), and a single agent-to-agent USDC transfer on Base costs cents in gas regardless of ticket size. Stripe's Machine Payments Protocol (MPP), co-authored with Tempo, sits in the same niche but extends to recurring payments and microsubscriptions.
​

Stablecoin settlement also relaxes the geography constraint. An agent paying for a Brazilian developer's API in São Paulo, a Singapore data feed, and a French inference endpoint runs into three different bank-account-holders, three different fee schedules, and three different settlement windows on traditional rails. On stablecoin rails, the same agent runs three transfers across whichever chain each provider accepts, in seconds, against the same dollar token. As more agentic payment surfaces multi-chain, the stablecoin orchestration layer (which chain to settle on, how to source liquidity, how to net obligations) becomes part of the production stack rather than an afterthought. Cross-chain messaging and stablecoin developer tooling are the two adjacent surfaces every serious agent-payment system has to integrate.
​

Sources and methodology. Stablecoin supplies and TVL pulled from DeFiLlama on April 30, 2026. Protocol launch dates and product details verified against issuer press releases and official documentation linked inline. Transaction volumes and active-agent counts from Cryptonews' April 2026 reporting. Figures refresh quarterly.
​

Eco operates as a stablecoin execution network across 15 chains, sitting beneath agent-payment protocols to handle cross-chain routing, solver selection, and finality. When an agent pays in USDC on one chain to a counterparty whose treasury sits on another, the orchestration layer abstracts the bridge, swap, and settlement steps into a single execution surface.
​

As agentic commerce settles in stablecoins across more than one chain, the cross-chain orchestration layer becomes load-bearing. Eco operates as a stablecoin execution network across 15 chains, abstracting routing, solver selection, and finality so that an agent paying in USDC on Base can transact with a counterparty whose treasury sits on Solana, Arbitrum, or Tron without writing the bridging logic itself. Agent-payment protocols (x402, MPP, AP2) handle the authorization and the rails; the orchestration network handles where the dollars actually live and how they get there. For teams building agent payment flows, that division of labor is the difference between integrating one execution surface and stitching together a bridge, a swap aggregator, and a settlement layer per chain.
​