Todd Bishop at GeekWire put it about as plainly as anyone could: Meta pointed its new agent at Amazon.com without any agreement in place. Not a pilot program. Not a handshake. Just a bot showing up at the world’s largest store and starting to browse on behalf of users who had no idea what was being recorded.
Amazon’s response, reported on September 20, 2026, was to block it. Security and privacy concerns, according to Amazon, including the finding that Muse was storing customer data without proper identification. Meta had launched the agent earlier that same month.
I review agents for a living, and this one gets a hard mark against it. Not because Amazon won the argument on principle — I’ll get to why Amazon’s motives are not pure — but because shipping an agent that can’t stay logged into the biggest retail site on the internet is shipping a product with a hole in the middle of it.
What actually broke here
Strip away the corporate press-release energy and the failure is boring and technical. Muse was reportedly storing customer data without proper identification. That is not a philosophical dispute about the future of commerce. That is an agent operating without a verifiable identity, doing things on a user’s account, and keeping records of it in a way the site owner could not audit.
If you ran a store and someone walked in wearing a blank name tag, filled a cart using a customer’s wallet, and took notes on the way out, you would ask questions too. Amazon asked, and then stopped asking.
For anyone evaluating agentic shopping tools right now, this is the criterion that matters more than benchmark scores or demo videos. Can the agent prove who it is? Can the site it’s acting on verify that claim? Can the user see what the agent kept? Muse, by the account that got it blocked, failed the first one and that made the rest moot.
Amazon is not the hero of this story
Amazon’s position is that it has a policy against unauthorized AI agents, and blocking Muse is consistent with it. Fine. It is also consistent with Amazon’s interest in owning the layer between you and your purchase.
Amazon has reportedly blocked dozens of AI bots, with reporting citing 47. But its subsidiaries — Zappos, Shopbop, and Woot — have stayed open to AI crawlers. That is not the behavior of a company with a settled moral stance on automated agents. That is a company running experiments in the parts of its business where the downside is small, while keeping the main property locked.
So read Amazon’s security-and-privacy framing as accurate and self-serving at the same time. Muse really did store data it shouldn’t have. Amazon also really does not want Meta standing between its customers and its product pages, harvesting behavior along the way. Both things fit in the same sentence.
What this means if you were planning to use Muse
A shopping agent that cannot shop at Amazon is a shopping agent with a major asterisk. Whatever else Muse does well, the single destination most Americans would point it at first is now closed. That is a functional limitation, not a temporary inconvenience you should price in as “coming soon.”
My practical read for anyone testing agentic commerce tools:
- Ask where the agent is actually allowed to operate. Not where it technically can reach — where it has permission. Those are different lists, and vendors love to blur them.
- Ask what the agent stores and how it identifies itself. If a vendor can’t answer this clearly, you are the audit trail.
- Assume more blocks are coming. Amazon moved first and loudest. Other large retailers watched this happen and took notes.
- Treat launch-day agent demos as marketing. Muse shipped in early September and was blocked from a flagship site within weeks. The demo was not the product.
The uncomfortable part for Meta
Launching into Amazon without an agreement was a choice. Someone at Meta decided that asking for permission was slower than asking for forgiveness, and that the story would be about AI progress rather than about data handling. Instead the story is a headline that says a major retailer found Meta’s agent storing customer data improperly. That is a worse outcome than a delayed launch.
The broader fight over who controls agent access to commerce is going to run for years, and I don’t think retailers get to wall themselves off forever. Customers will want agents that work everywhere, and eventually the economics will push toward negotiated access with real identity standards attached.
But that future arrives through boring plumbing work: verifiable agent identity, disclosed data retention, permissioned access. Meta skipped the plumbing and went straight to the launch. Amazon handed it the bill.
🕒 Published: