The core problem is simple: as AI agents gain the ability to shop, book reservations, and navigate websites on their own, businesses have no reliable way to know if the agent represents an actual person or a coordinated bot operation. World's AgentKit sits at the verification checkpoint. It cryptographically proves a verified human authorized the agent and remains in control of what it can do.
Access to AgentKit now includes Claude Code, Codex, Cursor, Hermes, and OpenClaw. The setup chain is mechanical: get a verified World ID through the World App, connect proof of human via the ToolRouter interface, generate an API key, and link the agent. World built this on top of x402, a payment standard Coinbase created to let agents transact onchain without halting for human approval at every step.
The March launch included a live test. World dropped 500 "Human in the Loop" hats via Shopify to verified World ID holders only. Agents used the search tool Exa to find the storefront, navigated Shopify's checkout, applied a discount code, and closed the order—all capped at one hat per verified human. Creating fake accounts or extra agents didn't unlock extra eligibility. All 500 sold across the US, Germany, Japan, and the UK.
Why it matters: the constraint revealed something concrete about the mechanism. A single verified identity can't be split into multiple agent personas with different purchase rights. That one-to-one binding is the whole point. If an agent could mint new agents and each new agent could claim its own allocation, the cap collapses into theater. Unchained Crypto reported the mechanics held during the hat drop, but real stress arrives when agents handle higher-value transactions or when incentives tip harder toward fraud.
World frames AgentKit as the foundation layer for an "agent economy" where software transacts on behalf of people at scale. The long-term play is turning World ID into the mandatory checkpoint for agent legitimacy, the way credit scores gate consumer lending. That requires agents to become useful enough and common enough that websites demand proof of a real human behind every autonomous action. Right now, most agents are still tethered to prototypes and demos. The infrastructure is live; adoption is nascent.