AI Agents Are Running Into a Wall of Closed Websites

Introduction

AI agents are being marketed as software that can do more than answer questions. They can search the web, book a table, clean up an inbox, or purchase an item on a user’s behalf. Yet once these systems begin acting on the open internet, many websites treat them less like assistants and more like untrusted automation. Users have recently reported that Meta’s Muse, xAI’s Grok, OpenAI’s Dots, and Instinct are unable to access or complete tasks on some major services.

Key developments

• Platforms are setting explicit conditions. Yelp told TechCrunch that it permits AI agents through its data-licensing program. Delta Air Lines does not allow agents to book tickets for users, while JPMorgan Chase and Charles Schwab have confirmed that they block agents from accessing accounts because of security concerns. Users have also reported problems with services such as Cars.com and Amtrak.
• E-commerce is the main battleground. Amazon previously blocked Perplexity’s agentic shopping feature and sued the company. Perplexity argued that its agent was an extension of the user, while Amazon said third parties buying goods for customers should operate openly and respect a provider’s decision to participate or not. The dispute highlights a central question: is an agent a user-controlled assistant, or an unauthorized automated visitor?
• Existing anti-bot defenses can catch agents. Many sites are not targeting AI agents specifically. Their anti-scraping, fraud-prevention, and account-security tools are designed to detect unusual automated behavior. Agents may trigger those safeguards because they log in, navigate pages, fill out forms, and attempt checkout without behaving exactly like a person.
• Partnerships offer a different route. Meta has argued that rejecting a personal agent also means rejecting the customer behind it. Rather than fight every retailer, it has partnered with Walmart, Gap, Sephora, Best Buy, and others. Shopify has likewise opened its merchant network to agents.

Why it matters

The dispute shows that an agent’s biggest limitation may not be reasoning quality. It may be permission. A system can understand a webpage perfectly and still lack the right to log in, place an order, make a payment, or access a financial account. Banks and airlines have strong reasons to be cautious: responsibility for mistakes, fraud, and compensation must be clear before automated actions are accepted.

Agents may also disrupt the economics of the web. If they compress a shopping journey into a few automated steps, retailers could lose page views, advertising exposure, and control over customer relationships. Amazon’s opposition to agentic shopping reflects this commercial concern as much as a technical one.

For users, the result could be a less seamless web. More websites may demand verification steps, challenge suspicious sessions, or simply refuse agent access. The irony is that automation intended to make online tasks easier could create more obstacles for people who use it.

The practical future may therefore be a permissioned internet for agents. Standardized interfaces, explicit consent, licensing programs, and clear limits on actions could allow businesses to participate without surrendering control. AI agents will become dependable assistants only when identity, authorization, data use, and liability are settled alongside model capability.

Source: The Verge AI