Shopping agents collide with a web that doesn't recognize them: Meta pushes a standard against Amazon and Delta

By Zendoric · October 7, 2026.

The facts are concrete: Amazon is blocking Muse, Meta's shopping agent, citing data security concerns, as TechCrunch reports. Walmart, an official Muse partner since Meta's Connect conference in September, attributes other failures to human verification checks (the typical "I'm not a robot" buttons) that the agent fails to complete properly. Delta confirms it has no integration whatsoever with external agents and that any opening up would first require guaranteeing customer security. United points to its terms of use, which expressly prohibit robots and spiders without written permission. And users report suspended accounts on eBay, cut-off access on Yelp, and even Instinct being kicked out while managing a Gmail inbox.

The industry's response is an open standard backed by Meta, Walmart, Stripe, Sierra, Genesys, Rocket, NiCE and Decagon: an agent-to-business communication protocol designed for e-commerce that would let a site verify that an agent is acting on behalf of a real customer and is not malicious traffic in disguise. Cloudflare, which some observers suspect of disrupting Muse traffic though the company itself has not confirmed it, adds another layer to the board: it operates a marketplace where AI bots must pay to access data and is testing a browser designed for agents. Meanwhile, Meta is negotiating bilateral deals brand by brand and maintains a list of authorized "connectors", still incomplete relative to the partners it has already announced.

This matters because it brings to the surface a problem that had been brewing for some time beneath the advertising promise of the assistant that solves everything. Shopping agents are the most tangible consumer AI use case of 2026, the one that turns natural language into an action with a credit card involved. But that action runs through an anti-bot infrastructure built over years to stop precisely what an agent does: browse without a human, fill in forms by machine, generate traffic that cannot be identified as a person. The result is not just technical friction, it is a conflict of incentives. Amazon has no commercial reason to make life easier for an agent that buys on its platform but costs it control of the customer relationship, the behavioral data and the chance to sell advertising or recommendations on the way to the cart. The same applies to Delta or eBay: every agent that buys without going through their own interface is a source of data and influence the company stops controlling.

Our reading is that this confirms a thesis we have been arguing for a while: the next big AI battle will not be fought over model quality, but over who controls the "plumbing" of distribution and access. An open standard backed by Meta sounds good on paper, but Meta is not a neutral referee: it is a player with a direct interest in Muse working, and the standard it proposes will tend to favor the personal agent architecture it has already built. Amazon, Delta or Yelp, for their part, have no reason to sign up to rules designed by a direct competitor in commerce or advertising. It is reasonable to expect adoption to be slow, uneven and littered with bilateral deals —like the one Meta already has with Walmart— rather than a universal norm. The web will probably fragment into two categories: sites that sign agentic access agreements (and charge for it, as in Cloudflare's model) and sites that simply close the door, preferring to protect their direct customer relationship even if that means turning away legitimate automated purchases.

In the short term, the user is the one who pays the friction: purchases that fail without explanation, accounts suspended by mistake, the sense that the assistant promoted in the keynote does not work in practice. It is a real problem, not a minor detail, and it should not be played down. But in the medium term, if agent identity standards mature —and there is plenty of money and corporate interest pushing in that direction—, today's friction will be resolved much as card payments on the internet were: with verification protocols we now take for granted. The underlying promise remains intact: delegating routine tasks to an agent so we can keep the time and judgment that do matter to us is exactly the kind of operational abundance this technology points toward. The obstacle is not the model's capability, it is that last century's web still cannot read the intentions of whoever visits it.