When the Agent Knocks: Amazon Blocks Meta’s Muse and the Fight for Agent Commerce Begins

The first great battle of agentic commerce was not fought over payments technology. It was fought over a locked front door.

At Meta Connect 2026, Mark Zuckerberg confirmed what many platform strategists had privately predicted: Amazon has blocked Meta’s Muse AI agent from shopping on its site. The disclosure arrived in the same keynote where Meta laid out its most ambitious agent roadmap yet — computer control from a new Mac app, a visual embodiment model, a plan to put Muse at the center of its smart glasses, and a commerce stack wired to Stripe, Shopify Shop Pay, and an expanded PayPal integration. Meta is even considering charging a small fee on transactions Muse completes.

Stand back from the product details and the structure is clear. The agent wars have moved past “which model answers questions best.” The new contest is which agent is allowed to act on the open internet — and who captures the economics when it does. Amazon’s answer, for now, is: not yours, not on our store.

What Meta Actually Announced

The Muse phase unveiled at Connect is built for long-running, embodied work:

Muse Spark 1.3, a model built specifically for long-horizon agentic tasks, underpins the new phase.
An updated Mac app lets Muse drive and use the computer directly — the same computer-use capability category that coding-focused agents and browser harnesses have pioneered this year.
An embodiment model lets users converse with a visual character rather than type, and Muse is being positioned as the primary agent on Meta’s smart glasses.
Muse for Small Business connects Shopify, Stripe, QuickBooks, Slack, Notion, Dropbox, and Canva, with approval gates on any action that sends, publishes, or spends money.

The traction numbers tell a two-sided story. Muse passed 5.6 million installs and briefly topped the App Store — genuine consumer demand. But open rates still trail mainstream assistants, and analysts have already begun marking down long-term monetization expectations. Distribution is not the same as habit, and habit is not the same as a transaction business.

Why Amazon Locked the Door

To understand the blockade, follow what a shopping agent actually does to a retailer’s economics.

Amazon’s marketplace is not, at its core, a product warehouse. It is a discovery and recommendation engine with a checkout attached. The value — and the profit pool — lives in deciding what shoppers see: sponsored placements, recommendation slots, the curated ordering of search results, Prime-driven loyalty mechanics. Sellers pay for access to that discovery surface; Amazon’s advertising business is built on it.

A capable third-party agent flattens that surface. When Muse shops “for” a user, it reads Amazon’s catalog as a data layer, compares it against other merchants, applies the user’s own priorities, and executes the purchase through whatever payment rail Meta has connected. The agent — not Amazon — owns the recommendation, the comparison, and increasingly the customer relationship. Amazon is left supplying logistics for a transaction in which its most profitable machinery was bypassed entirely.

Add Meta’s contemplated transaction fee and the provocation sharpens: Meta would monetize demand that Amazon’s inventory and fulfillment satisfy, while Amazon’s own discovery business is cut out of the loop. Blocking the agent is the rational response of a platform owner protecting the layer where its margin lives.

The pattern will repeat. Any incumbent whose profit pool sits in curating human attention — search engines, travel aggregators, marketplaces, app stores — faces the same question when an agent shows up: welcome the traffic, or defend the discovery layer? Expect doors across the internet to start closing to third-party agents over the next twelve months, alongside technical countermeasures: bot defenses, terms-of-service enforcement, and agent-specific access fees.

The Counter-Strategy: Agent Commerce Around the Walled Gardens

Meta’s response is equally instructive. Rather than negotiate with Amazon, it is assembling a commerce stack that routes around walled-garden retailers:

Stripe and Shopify Shop Pay give Muse native checkout across the universe of independent merchants.
PayPal widens the rail to legacy retailers.
The pitch to retailers is direct and plausible: the agent brings incremental demand, and small and mid-sized merchants who lack Amazon’s discovery monopoly have every reason to welcome it.
Approval gates on spending actions address the trust problem that limits every consumer agent: users will grant an agent permission to buy only when they control the threshold.

This is the agent-commerce world splitting into two spheres: closed super-platforms that admit no agent but their own, and an open merchant web served by independent payment rails and third-party agents. Which sphere grows faster depends on a variable still being measured — whether consumers actually delegate purchases, or merely use agents to research and then buy manually as they always have.

The China Counter-Example and the Open-Agent Playbook

It is instructive to compare the Western standoff with how agent commerce has evolved in China this year. There, the agent-commerce conversation has run through infrastructure the platforms themselves opened: the WeChat Pay AI card work and agent-integrated commerce flows treat third-party agents as participants inside a shared payment and identity system rather than as intruders at the gate. Super-app architecture — payment, identity, and merchant access already standardized — made agent participation an extension of an existing rail rather than a land grab.

That is the context in which open agent frameworks like DeepSeek Harness and the broader DeepThink ecosystem matter. Harness has spent 2026 building exactly the plumbing agent commerce requires: plugin architectures, cross-platform compatibility layers, sandboxed execution, and cordis-style connectors that let agents call external services through stable interfaces. An open framework does not eliminate the platform standoff — no framework can force Amazon to open its door — but it does three things that change the game:

It lets merchants choose agents instead of agents choosing merchants. Any retailer can expose itself to the whole ecosystem of open agents through one integration rather than cutting a deal with Meta alone.
It makes the agent swappable. When the harness, not the consumer-tech giant, owns the execution layer, the customer relationship and transaction data don’t inevitably accrue to one American platform company.
It gives payment providers neutral ground. Rails like Stripe and PayPal can plug into a standard agent protocol rather than anointing a winner in a platform war.

The Western blockade and the open-framework playbook are, in this sense, two bets on the same end state. Both assume that agents will eventually execute a large share of commerce. They disagree on who gets to toll the bridge: the platform that owns the store, the platform that owns the agent, or an open protocol layer owned by no one.

What to Watch Next

Three signals will determine how this resolves:

Whether other major retailers follow Amazon’s block, or accept agents in exchange for access fees and data-sharing terms.
Whether Meta’s transaction-fee model produces real revenue — which depends on converting 5.6 million downloads into habitual delegated purchasing, the step consumer agents have failed at so far.
Whether open harness protocols gain merchant adoption fast enough to present a neutral alternative before the market tips to one or two closed agent platforms.

Amazon slamming the door on Muse is not the end of agent commerce. It is the moment the industry admitted how much money is at stake and began fighting over the front door. The winners will be the players — open frameworks and open payment rails very much included — that make it rational for the most merchants to open theirs.