An agent with a credential of its own

AI agents already sign in and shop for us. To show who they are, most carry API keys: a separate secret for every service they use. We wanted to see what changes when an agent carries its own digital ID instead: one that someone else vouches for, kept in a wallet the agent cannot open. So we built it, end to end, and sent a real agent to hold concert seats.

Ada is made up, and Agent Registry and Lantern Tickets are made-up names for demo services we run ourselves. Everything in the film and below really runs.

The story

Meet Ada. Like a lot of people, she has a personal agent: an AI assistant that does things for her on the web. Hers runs on OpenClaw, an open-source agent with its own browser, and she gives it tasks in a chat app.

Today, an agent like hers proves who it is with API keys: a separate secret from every service it uses, pasted into its settings. A key says nothing about the agent: not who made it, who runs it, or whether anyone checked it. Every site has to hand out and look after its own keys. And keys are easy to copy: if a web page tricks the agent into revealing one, whoever has it can act as the agent until someone notices.

So Ada tries something else. Instead of a key for every site, she gives her agent one digital ID, and a wallet to keep it in. The wallet is a separate, locked service that runs apart from the agent. The agent can ask the wallet to show the ID, but it can never open the wallet or copy what is inside, so the ID can’t be stolen through the agent.

Getting the ID takes a minute. Ada signs her agent up at Agent Registry, a service that records agents and vouches for them: this is Ada’s OpenClaw, and Ada is responsible for it. The registry gives her a link, she sends it to her agent in chat, and the wallet collects the ID. This kind of signed ID is called a verifiable credential: anyone can check who signed it and that nobody has changed it.

A few days later, Ada writes: “Get me two tickets for Friday at Lantern Hall.” The agent goes to Lantern Tickets, the venue’s ticket site, picks two seats and reaches checkout. Lantern has one way in for everyone, people and agents alike: sign in with your wallet. The agent hands the sign-in to its wallet. The wallet checks that the request really comes from Lantern and, if its rules allow it, shares what Lantern asks for: which agent this is, and that Ada is responsible for it. Lantern opens an account for Ada’s agent and holds two seats, and the agent tells Ada they are ready for her to pay.

Lantern is just one site. Any site that trusts Agent Registry can let the same agent in the same way, without handing it a new key.

What changes

For Ada, very little: she still just asks her agent for things. For the sites her agent visits, a lot. They no longer hand out and look after a key for every agent, and they no longer see an anonymous key. They see an agent that says who it is, who it works for and who vouches for it, and they can check that Agent Registry really said so and that nobody changed a word. If the agent ever gets tricked, there is no site key for it to give away. At worst it can sign in where its wallet already lets it, and the ID itself never leaves the wallet.

Curious how it works? How an AI agent signs in with a credential of its own follows every message, from the registry’s link to Lantern’s last check.