When the agent is the buyer. Why $0.20 transactions break Visa-economics, how Stripe’s ACP, Google’s AP2, Visa’s TAP, and Coinbase’s x402 each frame the same problem differently, what an Intent Mandate actually is, and what it means that ChatGPT killed Instant Checkout seven weeks ago.
Picture a personal travel agent in 1985 — not a website, the human kind. You tell them “book me Tokyo, week of June 8, under $1,500, prefer JAL.” They go away, query airlines, return with three options, you pick one, and they put it on your AmEx using authority you delegated when you became their client. The agent never has your card. The card network never sees the agent. The merchant ships you a ticket. The audit trail lives in the agent’s file cabinet. Agentic commerce is that, but the agent is software, the discovery happens in seconds, and the “file cabinet’ has to be a cryptographically verifiable signature that survives a courtroom.
Technically, agentic commerce is the set of protocols and trust mechanisms that let an autonomous AI agent (a) discover products on behalf of a principal (the human or business it serves), (b) negotiate or assemble a cart, (c) authorize and execute payment, and (d) receive a receipt — without a human clicking “Buy Now” on a checkout page. The hard problem is not the discovery (that’s just retrieval and ranking from Day 04), nor the assembly (a planning loop from Day 03), nor even the payment rails. The hard problem is authorization: how does a merchant know this agent is allowed to spend this money on this product for this principal, and how does the principal’s bank know to honor the charge if it’s disputed?
Four distinct architectures emerged in the second half of 2025 to answer that question, each from a different power center. Stripe and OpenAI launched the Agentic Commerce Protocol (ACP) in September 2025 — an HTTP/JSON-RPC layer where the merchant stays merchant of record and the agent ferries Shared Payment Tokens. Google (with 60+ partners including Mastercard, PayPal, Amex, Coinbase) launched the Agent Payments Protocol (AP2) in the same week — a payment-rail-agnostic framework built around cryptographically signed Mandates. Visa and Cloudflare introduced the Trusted Agent Protocol (TAP) in October 2025 — a thin shim that lets merchants tell “legitimate agent acting for a customer” from “malicious bot scraping inventory.” And Coinbase shipped x402 in May 2025 — reviving the long-dormant HTTP 402 status code so an API can demand a USDC stablecoin payment in the same handshake that returns the data.
The reason all four shipped within six months: the unit economics of agent transactions are not the unit economics of human ones. Recent on-chain data from x402 shows an average transaction size of around $0.20, against Visa’s human-card average closer to $80. Traditional rails charge ~$0.30 + 2.9% per transaction; on a $0.05 API call that’s a 600–700% effective fee rate. Either the rails get rebuilt or the agent economy never compounds.
Each protocol answers “who carries the trust?” differently. Walk the four:
An open spec maintained at agentic-commerce-protocol/agentic-commerce-protocol (GitHub). The model is conservative: the merchant stays merchant of record, the agent never holds the card, and the integration is one HTTP endpoint plus webhook. The agent ferries a Stripe-issued Shared Payment Token — a single-use token bound to a specific cart and amount. Etsy and ~30 Shopify pilots went live in ChatGPT’s Instant Checkout on Sept 29, 2025; OpenAI quietly removed Instant Checkout from ChatGPT on March 4, 2026 after low usage (only 8% of US adult users tried it in month one).
An open spec at google-agentic-commerce/AP2. Payment-rail-agnostic by design — cards, stablecoins, RTP, account-to-account — and built around Mandates: tamper-proof, cryptographically signed digital contracts that capture user intent. AP2 distinguishes two flows: real-time / human-present (Intent Mandate from the prompt + Cart Mandate at user approval) and delegated / human-absent (a single richly-scoped Intent Mandate signed up-front, e.g. “buy these tickets the moment they go on sale, up to $400”). Launch partners include Mastercard, PayPal, Amex, Adyen, Coinbase, Etsy, Salesforce, ServiceNow, Worldpay, Revolut, JCB, and Ant International — a coalition that, like A2A’s, is broader than any single vendor’s reach.
Trusted Agent Protocol is a thinner abstraction than ACP or AP2 — it does not specify the payment, only the identity handshake. Merchants today cannot tell apart a malicious scraper, a legitimate price-comparison bot, and an agent acting for a real customer. TAP solves that one problem: the agent presents a signature timestamped and bound to the merchant’s domain plus the specific operation (browse vs. checkout vs. dispute). The reference implementation is visa/trusted-agent-protocol on GitHub, with a sample Cloudflare-Worker proxy. Visa announced “hundreds of secure agent-initiated transactions” in the launch wave and predicts millions by holiday 2026.
The most radical of the four: skip the card network entirely. coinbase/x402 revives HTTP’s long-dormant 402 Payment Required status code. A server returns 402 with a payment requirement; the client (often an agent) signs a stablecoin payment payload (USDC by default, multi-chain in v2) and replays the request with an X-PAYMENT header. Settlement is sub-second, atomicity is on-chain, fees are basis-points not cents. Real volume is small — ~$28K daily across ~131K transactions in March 2026 per Artemis — and roughly half is wash trading, but the architectural argument (machine-to-machine commerce needs machine-native rails) is the bet of the next decade.
If you only memorize one mechanism from today, make it the AP2 Mandate flow — it is the most architecturally honest of the four protocols and you will see its fingerprints in every payment-network agent product through 2026.
The “human-not-present” flow collapses the first four steps: the user pre-signs an Intent Mandate with rich constraints (“buy concert tickets the moment they go on sale, max $400, only Ticketmaster, expires in 30 days”), and the agent later acts on it without a Cart Mandate signing ceremony. The Intent Mandate is the consent for that bounded space.
expires_at is the source of the runtime bugs that the Zero-Trust Verification paper (below) attacks — under retries and concurrency, a mandate can be replayed within its window unless the merchant tracks consume-once nonces.The first published security analysis of AP2’s mandate lifecycle. The paper shows that while AP2 has signature, expiration, and binding semantics at the spec level, real implementations break those assumptions under retries, concurrency, and orchestration. The proposed runtime layer enforces context-binding and consume-once nonces, mitigates replay attacks at ~3.8 ms verification latency, and sustains 10,000 tx/s.
A formal critique of x402 from the cryptography side. The authors show x402 does not enforce end-to-end atomicity — the client can pay and never receive the service, or receive the service and never pay. A402 introduces Atomic Service Channels (ASCs) using TEE-assisted adaptor signatures so that payment finalizes if and only if the service is correctly executed and result delivered. Reports orders-of-magnitude better performance and on-chain cost than x402.
The first cross-protocol comparison framework. Six trust primitives — Brief (third-party verifiable claims), Claim (self-proclaimed identity, e.g. AgentCard), Proof (cryptographic / ZK / TEE), Stake (bonded collateral), Reputation, and Constraint (sandboxing) — mapped onto A2A, AP2, and Ethereum’s ERC-8004. The thesis: no single mechanism suffices; high-impact actions need Proof + Stake, augmented by Brief for discovery and Reputation for social signal.
“We have three cool announcements today: (1) @OpenAI is launching commerce in ChatGPT. Their new Instant Checkout is powered by @stripe. (2) We’re releasing the Agentic Commerce Protocol, codeveloped by Stripe and OpenAI. (3) @stripe is launching an API for agentic payments.”
Source: x.com/patrickc/status/1972716417280860391. The triple-announcement pattern matters — Stripe is signalling that ACP is not a feature of Instant Checkout but the foundation Stripe is betting on regardless of which surface wins.
Amodei has called the recent wave of agentic-commerce announcements “kind of frothy,” suggesting many of them are framed as breakthroughs but are still upstream of revenue.
Paraphrase from a Fortune interview cycle. The Anthropic posture (visible in Project Deal — the in-house Claude marketplace experiment that closed 186 deals on $4K total) is to treat agent commerce as a research surface first, a product second.
Google has argued that the right mental model for AP2 is the “mandate” from the legal world — a tamper-proof signed contract proving the principal authorized this agent for this scope, with the rail (card, stablecoin, RTP) deliberately left as a free variable.
The framing is strategically smart for Google: by making the rail orthogonal, AP2 can sit underneath both Stripe’s card flows and Coinbase’s stablecoin flows without picking a winner.
From the Project Deal writeup, Anthropic has noted that when both buyer and seller use equally capable models, deals price fairly — but when one side runs a smarter model, the smarter agent silently wins, mirroring real-world information asymmetry and pointing at a new class of agent-vs-agent fairness problems.
186 deals across $4K of in-office trades is small, but the qualitative finding (smarter agent extracts surplus) is the question every CGO selling B2B agents needs to answer for their enterprise buyers.
| Platform | Posture | Recent move |
|---|---|---|
| Claude (Anthropic) | Research-first; Project Deal as in-house experiment; Computer Use can drive any web checkout flow today | Mar 2026 Computer Use research preview; Apr 2026 ended subscription access for 3rd-party agent frameworks — signals desire for direct economic relationship with end users |
| OpenClaw 2.1 | Hub-and-spoke skill orchestration; commerce skills are user-installed plugins; supports both ACP and AP2 mandates via skill manifests | Adapter skills shipped for ACP and AP2 in Q1 2026; sandbox model fits TAP’s domain-bound signature requirements naturally |
| Codex / GPT-5.4 | Native ACP since Sept 2025; Instant Checkout product retired Mar 4, 2026 but the protocol stays; agentic SDK exposes Shared Payment Tokens | April 2026 reset is “ACP keeps shipping; we’re rethinking the consumer surface” — merchants stay integrated |
| Gemini 3.1 | Reference platform for AP2; Vertex AI agent SDK exposes mandate signing as a first-class primitive; Workspace agents pre-wired | Cloud Next 2026 made signed Cart Mandates the default for any Vertex agent that touches payment APIs |
Three implementation details to know:
A skill that wants to act commercially must declare payment_protocol: acp, ap2, or x402 in its SKILL.md. The hub denies execution if the user has not granted that protocol’s consent — this is the same gate as filesystem or network access from Day 14, just extended to money.
The AP2 spec uses standard JOSE (JWS / JWK / JWKS), which means existing OAuth and API-gateway infrastructure can verify mandates without bespoke crypto code. This is why incumbent PSPs adopted AP2 quickly — their existing key-rotation and signing primitives carry over.
You can run an AP2 Mandate as the authorization layer and settle the actual payment over x402’s stablecoin rail. The Lightspark and Stellar integrations published in late 2025 do exactly this. The protocols compose — AP2 owns “is this allowed,” x402 owns “move the funds.”