Three dev.to posts map the agentic commerce stack and build a working MCP shopping agent

A three-part tour of agent-driven purchasing

On November 14, a writer publishing as brudy_dev24 on dev.to released three related pieces: a plain-English guide to agentic commerce, a map of the five protocols behind it, and a step-by-step tutorial that builds a small but functional shopping agent. Together they offer both a vocabulary for the emerging stack and runnable code to start from.

Agentic commerce, as the guide defines it, is an AI agent completing part or all of a purchase on someone's behalf — research, selection and payment — rather than a human clicking through checkout. The guide distinguishes consumer-facing agents (assistants buying on request) from business-facing agents doing continuous procurement, and argues that most current infrastructure work is about making higher-value, unattended purchases safe enough to trust.

How an agent actually buys

According to the dev.to guide, an agent purchase is a chain of machine-to-machine steps, not a human filling in a form: the user states intent with constraints such as budget, the agent discovers products through structured machine-readable feeds, selects an option, proves it is authorized to spend, pays, and then confronts the hardest part — verifying that the deliverable matched the agreement and recovering if the agent process dies mid-transaction.

The guide cites a stack of market signals: Forrester puts AI agents in 11% of US online transactions over $75; Shopify says AI-driven traffic tripled year over year in Q2 2026; Morgan Stanley estimates agentic shoppers could represent $190–385 billion in US e-commerce by 2030; Bain projects 15–25% of US sales, or $300–500 billion; McKinsey sizes the global opportunity at $3–5 trillion; and Juniper Research sees transaction value growing from $8 billion in 2026 to $3.5 trillion by 2031. The guide itself cautions that these forecasts use different methodologies and should be read as directional.

Five protocols, five different jobs

The stack explainer maps each protocol to one layer and, usefully, to what it does not settle:

• MCP (Anthropic, November 2024) standardizes tool access — often described as "the USB-C for AI" — letting agents call merchant APIs. It says nothing about what happens after a tool call creates a business obligation.
• x402 (Coinbase, May 2025) makes payment a native HTTP exchange: a server answers with HTTP 402 plus payment terms, the agent pays in USDC on Base, and the resource is released. The article reports over 165 million agent transactions processed by mid-2026 and backing from a foundation including Visa, Mastercard and Stripe. It proves money moved, not that the deliverable was correct.
• AP2 (Google, September 2025, with 60+ partners and later donated to the FIDO Alliance) provides cryptographically signed Intent, Cart and Payment Mandates. It captures user authorization but is not a full agreement record.
• ACP (OpenAI, Stripe and Meta, September 2025) defines the checkout flow between agent and merchant and powers Instant Checkout in ChatGPT. It is largely silent on post-checkout failures.
• EAS (Ethereum Attestation Service) provides signed, verifiable attestations on- or off-chain to anchor evidence — but a signature on a statement is not proof the commerce terms were met.

A runnable shopping agent

The tutorial builds a Node.js MCP server exposing two tools — a product search over a small catalog and a checkout-adjacent tool that records a commitment before any payment — plus a client that lets an MCP-capable model call them. The emphasis is on three properties the author says separate a trustworthy agent from a demo: durability, so the record survives a crashed process and a replacement agent can continue; authorization, bound to the buyer's verified signature so a substitute agent cannot inherit spending authority; and idempotency, so a retried request returns the original result instead of double-charging.

The gap the pieces keep pointing at

The through-line across all three posts is a missing layer: a durable, structured record of parties, terms, acceptance criteria, evidence requirements and recovery policy that exists before a transaction and outlives it. One caveat worth flagging: the articles repeatedly link to SAAX, a protocol the author documents, as an implementation of that layer, so the framing favors it. The argument itself is straightforward regardless — authorization proves permission and payment execution proves money moved, but neither preserves the agreement nor verifies fulfillment.

Why it matters

Agents are already transacting at meaningful scale, if the cited Forrester and Shopify figures hold, and merchants plus developers are building for them now. The protocol map matters because it tells builders which layer they are actually solving — and prevents duplicating work that MCP, x402, AP2 or ACP already covers. The tutorial matters because it lowers the entry cost from concept to working code. And the durability question the pieces raise is the one that will decide whether agentic commerce stays a convenience for $50 purchases or becomes trustworthy enough for the $5,000 contract.