Sam Altman's World tests AgentKit to link AI agents to iris scans

World, Sam Altman's project, has begun limited beta testing of AgentKit — a tool intended to link an AI agent's actions to a specific person through an iris scan. The idea is simple: if an agent acts on someone's behalf, the system should understand exactly who is behind that action.

What World is launching

World, known as WorldCoin until the end of 2024, announced the start of a limited beta test of AgentKit.

For the project, this is a logical next step: World has long been building infrastructure around digital identity, and is now trying to apply it to the fast-growing market for AI agents.

Instead of an abstract conversation about “trusted AI,” the company is offering a concrete framework: there is an agent, there is a verified owner, and there should be a verifiable link between them.

In practice, this is not just about a new interface for a bot. AgentKit is presented as a mechanism that will make it possible to directly tie an agent's actions to a person who has confirmed their identity through an iris scan. In World's logic, this should reduce the risk of abuse when autonomous systems masquerade as real users, create fake accounts, inflate activity, or perform actions without a clear level of accountability.

Why it is needed

The market for AI agents is moving quickly from demonstrations to real-world scenarios: bookings, purchases, account management, applications, and internal business processes. As soon as an agent gets the right to do something on a person's behalf, a basic question immediately arises: who is responsible for its actions, and how can that be proven to a service, a bank, a marketplace, or a regulator.

This is where World is trying to position itself with AgentKit.

• Distinguish personal agents from anonymous bot networks
• Tie access rights to a verified owner
• Simplify auditing of actions if an agent places an order or works with an account
• Increase the trust of platforms that accept automated requests on a user's behalf

In essence, World is offering an identification layer for the agent economy. If the idea takes off, platforms will be able not just to see that a bot has reached them, but to understand that a specific verified person stands behind it. For consumer services, this could become an argument in favor of allowing AI agents into more sensitive operations, and for the corporate market, a foundation for access policies, limits, and the subsequent review of disputed actions.

Where the risks appear

The strength of World's approach also looks like its most controversial point. The more tightly an AI agent's actions are linked to a person's biometric identity, the higher the cost of an error, a leak, or opaque access management. Users and regulators will almost certainly not be satisfied with a general idea of security: they need answers to questions about data storage, the ability to revoke permissions, access delegation, and exactly what traces remain after the agent has operated.

“build trust in the system”

The problem is also that trust cannot be created by an iris scan alone. Even if raw biometric data is not directly involved in every operation, the very chain of “person — verified identity — autonomous agent” becomes critical infrastructure. If such a setup fails, the user risks losing not only convenience, but also control over who acts on their behalf and how. And for a global product, this automatically means a tougher conversation about compliance across different countries.

What it means

World is trying to occupy an important niche between the AI agent boom and the crisis of trust around it. If AgentKit shows that agents can be made not only useful but also accountable, the market will get a new standard of digital identity. If not, the very idea of tying autonomous actions to biometrics will remain too sensitive even against the backdrop of the broader AI hype.