Cloudflare says websites pay for almost all 14 AI agent problems

On November 5, 2026, Cloudflare published a long paper about AI agents that work for persons on the web. It says that a website sees only the agent and not the person. When a user disputes an order or reads a page with no ad view, the site pays. In almost all of the 14 problems, the site has the primary cost, Cloudflare says.

How it works

Cloudflare calls an agent that operates away from the device of the user an intermediary. It is between the user and the website. Its rule says that the business relationship and the tasks to keep data safe do not change when an agent is between them.

The paper also gives a scale with 4 levels of user identity. It starts the Trusted Agent Delegation coalition for work with standards groups.

The numbers

In one example, a user lets an agent buy an item and then disputes the purchase. The site cannot use its usual data, such as the device and IP address, as the purchase comes from the infrastructure of the agent. If the site wins, the dispute has a cost too.

What is not known

The paper gives no figure for the traffic of agents at this time. It gives no date for when agents send more traffic than persons.

Insight: Cloudflare can write the rules for agent identity

On November 6, Cloudflare reported revenue of $782.5 million, 39% more than one year before. Its chief executive said that some customers want to keep agents out. On November 5, it asked for open standards on agent identity. This reads as a race to set the rules for agent identity, where the company that operates the gate can also write the standard. Leaders who operate sites must work with the standards groups from the start.