United won't let AI agents book flights and eBay appears to block them. Don't build client workflows on sites that can shut agents out
Two big platforms closed the door on outside AI agents this week. If your automation browses someone else's website, that door is part of your system.
What happened
United says it will not let personal AI assistants book tickets, The Verge reports, citing The Wall Street Journal. Agents can still show users United flights but cannot complete the transaction. United worries agents could expose customer information, make purchases people did not intend, and skirt federal rules such as having travelers attest their bags hold no explosives. CEO Scott Kirby said he does not want to create another middleman. United said it could reconsider as agents advance.
Separately, a reader told The Verge that Meta's Muse agent, which had been pulling daily eBay watch listings for weeks, was suddenly blocked. The Verge reproduced it. eBay already bans "buy-for-me agents" that place orders without human review, and a spokesperson said it has restricted agent access in certain cases while talking with potential partners.
The picture across retail is mixed. Amazon blocks external agents entirely, Target blocks them in some cases, and Shopify has opened a catalog of a billion listings for agents to buy from.
My take
I get asked to build agents that check supplier prices, monitor listings or book things on a client's behalf. These stories are a reminder of where that kind of build breaks: not in the model, but at the site's front door.
How I scope these now:
- Use the official API or data feed first. Browsing is the fallback, not the plan.
- If browsing is the only route, read the site's terms and expect a block. Build a check that alerts a human when the agent gets refused, instead of failing quietly.
- Keep the final step with a person for anything that spends money or makes a commitment. United and eBay are both drawing that same line.
- For clients who sell online, decide your own policy. Agents will arrive. A clean product feed or a booking API is a better welcome than a scraper fighting your checkout.
More posts
- Open OCR models now read 1,000 pages for under $0.10, and better serving saved more than switching modelsNov 20, 2026
- Gemini 4 Neon costs $3 in and $15 out per million tokens and tops Zapier's AutomationBench at 56.3%, though it uses more output tokens per taskNov 20, 2026
- Gemini API's Tool Retrieval loads tool schemas only when the model asks for them, so agents with hundreds of tools send fewer input tokensNov 20, 2026
