A quiet revolution is underway in retail, and it’s happening in the fine print. As AI agents start to act and shop on our behalf, major platforms are deciding whether to treat those agents as customers or intruders.

Over the past year, Amazon has repeatedly updated its robots.txt rules to bar AI-related crawlers and agents from accessing its storefront data. eBay updated its user agreement to explicitly ban “buy-for-me agents.” And in Europe, the European Commission opened an antitrust investigation into Meta over WhatsApp policy changes that restrict third-party AI chatbots’ access to the platform.

These fights aren’t just another spat between platforms and scrappers. They’re an early skirmish in a much bigger battle over the shape of agentic commerce.

The strongest case for allowing agents to act and buy on your behalf comes from Andrey Fradkin and Alex Imas. As they argued in “Why Can’t Your AI Agent Book a Flight?,” AI agents could revolutionize commerce but are stifled by two major impediments.

The first hurdle is technical. The internet is designed for humans, not machines, which makes simple tasks like clicking through a CAPTCHA or interacting with a dynamic page difficult for an agent. But private interests are already solving this problem. Former Twitter CEO Parag Agrawal has raised $100 million to build agent-friendly infrastructure.

The more intractable problem is the law. Companies are changing their Terms of Service agreements to prohibit automated tools, and precedents such as Facebook v. Power Ventures mean that platforms can revoke access at will.

To unlock new sources of value and competition, Fradkin and Imas make the case that “users have the right to deploy AI agents” assuming that the agent

• operates through the user’s own browser and credentials;
• acts only at the user’s direction;
• identifies itself as an AI agent operating on behalf of a specific user; and
• does not engage in data harvesting beyond what’s needed for the transaction.

There is a logic in this line of argument. If you can hire a human personal shopper or a travel agent, why not an AI one?

The strongest counter against this idea is that it would require the government to override private contracts and mandate that platforms accept AI agents. That’s not a small tweak. Companies would have to re-engineer their systems to accommodate third-party software before there are even settled standards.

But the more practical objection is that agents change the scale and scope of interaction. A human shopper is slow, rate-limited by biology. But an agent can be replicated and run in parallel. So even if an agent acts on the behalf of a user, the speed and volume with which it interacts breaks the assumptions under which today’s commerce systems were built. And one of the assumptions is that travel agents have fiduciary responsibilities.

Fradkin and Imas assume we can solve this with “reasonable security requirements,” but we’re at the very beginning of understanding these risks. As Tomašev et al explain in “Distributional AI Safety,” “prototyping and evaluating different steerable market designs” starting with “narrow, well-defined, and controllable use cases.” Mandating agent access before we have robust solutions is premature.

Instead, the market is tending towards three governance models:

• Hard exclusion — eBay is the clearest example of this. Bots simply cannot act on a user’s behalf.
• Guardrails — Shopify requires that “buy-for-me” agents include a human review step and directs developers to integrate Shopify’s checkout technology into their tools.
• Protocols — The biggest AI players, like Google, Anthropic, and OpenAI, are developing their own protocols so agents and merchants can easily talk with each other.

It’s worth noting that most retailers aren’t taking eBay’s approach of hard exclusion. Indeed, when Google just released Universal Commerce Protocol they were explicit that the standards were “co-developed with industry leaders including Shopify, Etsy, Wayfair, Target and Walmart” and are “compatible with existing industry protocols like Agent2Agent (A2A), Agent Payments Protocol (AP2) and Model Context Protocol (MCP).”

Like interoperability and net neutrality before it, the fight over agentic protocols will become a major policy issue in the near future, which economist John Horton highlighted on X, saying “I do wish people who are really energized about platform competition/market power would pay attention to this—before norms/practices settle, there’s an opportunity to create pro-competitive, pro-consumer policy around agentic commerce.”

The right move, however, isn’t to freeze the ecosystem by blessing one protocol or another. We don’t yet know which protocol will prove most robust, or whether the winning architecture will be a hybrid that borrows from each. Besides, as agents become more capable they could translate, route, and negotiate across multiple protocols at once, choosing the path on a transaction-by-transaction basis. In this highly contested market, policy should aim less at picking winners and more at preserving the conditions for open experimentation.