The Verge reported on 17 October that the web is closing to third-party AI agents faster than the agents are arriving. Amazon has blocked Meta's Muse since 20 September. Yelp permits only agents it has licensed. Delta says it has no agreements with third-party agents at all. None of the three is an AI company, and none framed the decision as being about AI quality.
Read alongside Schwab and JPMorgan blocking agents from moving money, the pattern is consistent across four unrelated sectors - retail, reviews, travel and finance - within roughly a month. The common thread is not that agents perform badly. It is that an agent arriving with a customer's credentials is indistinguishable from an attacker arriving with the same credentials, and the site carries the liability either way.
This is the gap between agentic demos and agentic products. A demo books a flight on a site that has not thought about it yet. A product needs the site to have decided, in advance and in writing, that the agent is allowed. Yelp's licensing approach and Schwab's read-only MCP plan are both early versions of that contract. Until more sites publish one, the practical ceiling on consumer agents is set by whoever says no, not by the model.