AI agents appear, who is behind them?

AI agents are becoming active participants in the Internet economy

They browse websites, manage accounts, and complete purchases on behalf of users. As capabilities continue to increase, the platform faces a new problem: When AI agents appear, how to confirm that there is a real human being behind them?

For years, platforms have relied on Captcha, rate limits, phone verification and Know Your Customer (KYC) checks to distinguish legitimate users from automated abuse. These tools are designed for an Internet where automation is mainly seen as needing to be blocked.

This assumption is disintegrating. Users increasingly want software to act on their behalf, and the scale of the relevant data is too large to ignore. McKinsey estimates that by 2030, AI agents could facilitate $3 trillion to $5 trillion in global commercial transactions. Bain's forecast for the U.S. market is more conservative, at US$300 billion to US$500 billion, accounting for approximately 15% to 25% of e-commerce. The difference between these two numbers reflects different understandings of the definition of "proxy" and is not a definite number, but no matter which prediction, it means that the platform is facing actual pressure.

A more thorny question than "humans or robots"

The past binary division of "humans are good, robots are bad" does not apply here. A legitimate AI agent can generate automated requests and complete transactions without requiring a human to click on every button-which looks a lot like abuse from a platform perspective.

Platforms face several imperfect choices: either block automation, making it inconvenient for authorized users, relax restrictions, encouraging abuse, or rely on strict verification that requires users to provide large amounts of files that many users are reluctant to do. There is currently no perfect solution.

One idea that some developers are gradually accepting is that the real question is often not "Who is this person?" It's "Is there a real and unique human being behind this agent?" This is a narrower issue and requires less personal information.

Different approaches to "human proof"

Some projects are building personality proof infrastructure, from biometric verification to social graph verification to hardware-based credentials. Each method makes a trade-off between convenience, privacy and the ability to resist cheating on a large scale.

A prominent example is the World Project, whose World ID protocol is described in its developer documentation as a way to protect privacy and prove the authenticity and uniqueness of a person's online identity without sharing personal information. World said zero-knowledge proof allows services to confirm a valid World ID without revealing the ID itself or correlating activities between different applications.

World has extended this functionality into the agent realm with AgentKit. AgentKit allows verified users to delegate their World ID to AI agents, allowing the agents to carry encrypted evidence that there are real humans behind them. The tool is built on Coinbase's x402 payment protocol and allows websites to ask for proof of a unique human being before granting an agent access. Since then, World has expanded its integration to platforms such as Browserbase, Exa, Okta, Shopify and Vercel.

Whether this model will eventually become the standard or one of several competing options is still inconclusive and likely depends on how much the platform gradually adopts it.

Unsolved Issues

In theory, an effective human certification layer could reduce false accounts and abuse without requiring every user to pass cumbersome KYC document reviews, while preventing users from handing over unnecessary personal information.

However, achieving this prospect requires more than just good design. Outstanding questions include how these systems will behave at large scale, how to handle running multiple legitimate agents, and how long it will take for malicious actors to adapt once a method is widely deployed. Such schemes need to prove that they are more difficult to deceive than the inspection methods they replace.

A broader shift still seems to be real: As agents take on more and more tasks for humans, platforms will need some way to define responsibility beyond the simple dichotomy of "human" or "robot." World's solution is a strong candidate for building this mechanism, while other competitive standards and innovative ideas are still emerging.