The Agentic Wall: Why LLM Power is Failing at the Front Door

The industry is currently obsessed with 'agentic' AI—the shift from chatbots that merely answer questions to agents that execute tasks. As TechCrunch first reported, we see this in the rollout of Meta’s Muse, Instinct, and ChatGPT’s Dots. These tools are designed to handle the friction of modern life: booking flights, ordering groceries, and making dinner reservations without requiring the user to be a technical expert capable of configuring tools like OpenClaw.

But as a practitioner, I see a glaring disconnect. The intelligence of the model is irrelevant if the destination website refuses to let the agent in. We are hitting a wall of legacy security and intentional corporate blocking that proves the web is not yet ready for autonomous agents.

According to reporting from TechCrunch, this friction is manifesting in two distinct ways: intentional blocks and accidental security triggers.

On the intentional side, companies are treating AI agents as unauthorized intruders. Amazon has recently blocked Meta’s Muse from browsing its product catalog or making purchases. Yelp has taken a more transactional approach, telling TechCrunch that it prohibits non-human traffic unless the agent has paid for access via a data licensing program. This means agents attempting to book tables or join waitlists will likely fail without a formal partnership.

Then there is the airline sector, where security concerns are paramount. Delta’s general manager of Global Communications, Heena Chavda, told TechCrunch that the company is taking steps to prevent unauthorized automated automated activity and currently has no integration allowing third-party agents to book flights. United, meanwhile, pointed to Terms of Use that prohibit the use of robots or automatic devices to monitor the site without written permission.

However, some of the most frustrating failures are accidental. Walmart is actually partnered with Muse—a collaboration announced at Meta’s Connect developer conference in September—yet users are still reporting failures. TechCrunch notes that Walmart’s site uses human-verification buttons to stop spam; if an agent triggers this check and fails, it is booted. In some cases, these security measures are so aggressive they impact humans, as evidenced by a social media post from Kevin Wang who claimed he could not sign in to his own account.

Other brands mentioned by early adopters on social media as blocking agents include Zillow, Pizza Hut, and Adidas. Some users report more severe consequences; TechCrunch notes reports of eBay suspending accounts due to agentic AI use, though eBay told the outlet it only restricts unauthorized actions like model training and automated scraping.

To solve this, a coalition including Meta, Walmart, Stripe, Sierra, Genesys, Rocket, NiCE, and Decagon is developing an open standard for agent-to-agent communication. Meta stated in a blog post that this protocol will focus on online commerce to help businesses distinguish between 'good bots' acting for users and malicious ones.

Until these standards are adopted, the 'agentic future' remains a theoretical exercise. The models are capable, but the gatekeepers are still using a 2010s security playbook.