Meta’s Muse AI agent has lost the ability to retrieve eBay listings, exposing a critical divide between an assistant's reasoning capabilities and its actual marketplace access. While Muse previously used browser-based scraping or API connections to aggregate product data, eBay has implemented selective restrictions to block unauthorized data collection. This highlights that "agentic commerce" is not a universal feature; functionality depends entirely on whether a platform permits the agent to access its inventory. Users should treat shopping assistants as fragile dependencies, as marketplace access rules can change, rendering previously reliable automated shopping workflows suddenly nonfunctional.
Meta’s Muse AI agent has stopped retrieving eBay listings in cases reproduced by The Verge, putting a practical limit on its shopping capabilities. In a report published on November 10, 2026, the outlet described failed requests for listing links and product details. eBay confirmed that it restricts some third-party agent access, while Meta declined to comment.
The immediate consequence is straightforward: a shopping assistant that previously assembled information from a marketplace may suddenly lose that source. Better reasoning, more detailed prompts and permission from the user cannot guarantee that the marketplace will let the assistant in.
This is a platform-access conflict with consequences for anyone evaluating agentic commerce, where software searches, compares or buys products on a person’s behalf. The important distinction is between what an agent can technically attempt and what a marketplace permits it to do.
What Stopped Working in Muse
A reader told The Verge that Muse had spent several weeks producing a daily digest of watches from sites including eBay. The digest included listing links, prices, shipping details and notes about condition or style.
Over the weekend before the report, Muse stopped returning eBay links and details, telling the user that it was blocked. The Verge reproduced the problem with other searches. In some cases, Muse directed users to outside websites that mirrored eBay listings.
That fallback matters. Finding a reference to an item elsewhere is different from retrieving its current marketplace listing. For a watch purchase, price, availability, condition and shipping information all affect whether an apparent match is useful. A mirrored page may provide a lead, but readers should verify those details on the original listing before acting.
The evidence has limits: The Verge did not publish a full test protocol, and eBay did not directly confirm that it had recently introduced a Muse-specific block. The restriction’s scope and duration remain unclear. The supported finding is that Muse failed to retrieve eBay information in reproduced cases, alongside eBay’s confirmation of selective restrictions.
eBay Separates Marketplace Access From Purchase Approval
eBay spokesperson Sybille St. Arromand told The Verge:
“We have restricted agent access in certain cases, including unauthorized scraping and training on our data,”
She also said eBay was discussing potential partnerships and evaluating ways third-party agents could help customers discover its inventory. eBay’s position leaves room for approved relationships while restricting other forms of access. The statement did not identify which activity triggered the reported Muse failures or accuse Meta of using eBay listings for training.
Separately, The Verge reported that eBay had already updated its user agreement to prohibit buying agents and large-language-model-driven bots that place orders without human review.
These are two different boundaries:
Access rules govern how an agent retrieves and uses marketplace information.
Purchase rules govern whether and how an agent can place an order.
A user can approve a purchase while the agent still lacks permission to collect the listing information needed to prepare it. Equally, an agent might be allowed to browse products without being allowed to buy them autonomously.
That distinction is particularly relevant to Muse. At its September launch, Meta said the assistant would seek approval before completing purchases, according to Engadget. Human confirmation addresses one part of the transaction. It does not, by itself, settle the marketplace’s rules for automated access.
My view is that shopping-agent evaluations need to treat these as separate requirements. A purchase-approval screen can be a useful safeguard, but it tells readers little about whether the agent has a reliable, permitted connection to the store.
Why Muse’s Connection Method Matters
Meta introduced Muse on September 8, 2026, as a personal agent for tasks including shopping, email and travel planning. Launch coverage described several ways it could reach external services.
Public API connections that Muse can set up using credentials supplied by the user.
Browser access when an API is unavailable.
An API is a structured interface through which software requests data or performs actions. Browser access instead lets an agent interact with a website through its web interface. These routes have different dependencies, even when they produce similar-looking answers.
An API connection still depends on the service’s authorization and usage rules. Browser access depends on the website continuing to accept the agent’s activity. Neither route gives an assistant unrestricted rights to another company’s inventory.
The available reporting does not identify the exact route Muse used for the failed eBay requests or the technical method behind the restriction. It would be speculation to attribute the failures to a particular bot filter, account rule or network block.
The practical implication is nevertheless clear: the connection is part of the product’s capability. An agent can understand a request perfectly and still fail because its route to the necessary information is unavailable.
For a recurring watch digest, that creates a maintenance problem. A successful setup last month is insufficient evidence that tomorrow’s report will cover the same marketplaces. I’d treat an automated shopping connection as a dependency that needs monitoring, especially when the task runs in the background.
Marketplaces Are Taking Different Approaches
The Verge’s report describes a fragmented landscape, with platforms making different decisions about third-party agents. Its November 10 overview provides this comparison:
Selective restrictions, with potential partnerships under discussion
Users cannot assume every third-party agent has the same access
Amazon
Blocks external agents from browsing or shopping
General web-shopping capability does not guarantee Amazon coverage
Target
Restrictions permit some browsing attempts while blocking others
A successful request may not predict reliable ongoing access
Shopify
Opened its catalog to support agent-driven purchasing
Catalog access offers a more explicit route for agent commerce
These are reported platform stances, not an independent compatibility audit of every agent, merchant or transaction.
For readers comparing assistants, a broad claim that a product can shop online is therefore incomplete. Useful coverage depends on the particular marketplace, the connection method and the action being attempted.
The same fragmentation has a competitive implication for the AI industry. If a marketplace selectively admits partners, commercial relationships could influence which assistants offer dependable access. That is an inference from eBay’s stated partnership approach, not evidence of a particular deal or preferred partner.
Model capability remains important. Access agreements can still determine whether that capability reaches the inventory a buyer actually wants.
Early Commerce Numbers Need Context
The Verge also reported modest third-party-agent usage figures from other services:
Uber: four rides fulfilled through third-party agents in the third quarter.
Uber Eats: around 500 orders during the week of Halloween.
DoorDash: several hundred orders placed through third-party agents.
For scale, DoorDash reported 993 million total orders in its latest quarter. However, the article did not provide a matching time window for its agent-order figure, so calculating an agent share from those numbers would be misleading.
These are company-reported figures relayed by The Verge, and they concern rides and food delivery, not eBay purchases. They cannot establish the size of the entire agentic-commerce market.
They do offer a useful counterweight to sweeping claims about assistants transforming everyday transactions. Some large services are reporting small volumes through third-party agents, even as companies build products around the promise of delegated shopping.
For buyers today, dependable coverage of a few relevant stores may be more valuable than a long list of actions an assistant can theoretically attempt.
How to Evaluate a Shopping Agent’s Access
The eBay-Muse case suggests a practical evaluation method: check permission, coverage and freshness separately. A polished answer can conceal a weakness in any one of them.
Verify the Marketplace and the Action
Ask whether the assistant supports the specific store and task you need. Searching listings, reading shipping details, adding an item to a cart and completing payment are separate operations.
Look for documented support or an approved integration where available. Review the marketplace’s current automation terms, particularly before enabling purchases. One successful search should not be treated as permission for every subsequent action.
For an integration project, this distinction belongs in the requirements. “Supports shopping” is too vague to describe a workflow that depends on retrieving listings and completing checkout at a particular marketplace.
Check the Original Listing
A useful sample request should return links that open to the intended products, with enough information to verify price, condition, availability and delivery terms.
If an assistant supplies a mirrored page, ask it to identify that source clearly. Then check the original marketplace. In the documented Muse cases, outside mirrors were a fallback after direct retrieval failed; they should not silently substitute for current eBay information.
For secondhand goods, I would put particular weight on this check. Individual listings can differ in condition and shipping terms, so a convincing product summary is insufficient on its own.
Make Missing Coverage Visible
For recurring searches, ask the agent to name the marketplaces it successfully checked and flag those it could not reach. Otherwise, an apparently complete daily digest may simply contain results from fewer sources.
A useful instruction would require the assistant to distinguish direct marketplace results from outside references and report access failures explicitly. This does not overcome a block, but it makes the resulting gap visible.
Keep a manual route available for important purchases. When a marketplace becomes inaccessible, the immediate fallback may be opening its site yourself. Another automated route would still need to respect the platform’s rules.
The central lesson from eBay and Muse is about reliability: access belongs in the assessment of an agent from the beginning. An assistant’s ability to understand a shopping task and its permission to carry that task out are separate dependencies.
Frequently Asked Questions
3 questions
1
Is eBay Blocking Every AI Agent?
eBay says it selectively restricts third-party access. It did not confirm a universal ban or a new Muse-specific block, although The Verge reproduced Muse’s inability to retrieve eBay listings.
2
What Do eBay’s Rules Say About Automated Purchases?
According to The Verge, eBay’s user agreement prohibits buying agents and LLM-driven bots that place orders without human review. Separate access restrictions can affect browsing and data collection before checkout.
3
How Does Muse Connect to Other Services?
Launch reporting describes built-in connectors, public API connections using user-provided credentials, and browser access when no API is available. The reporting does not identify which route Muse used in the failed eBay requests.
