Every agent needs a home and an address

Last weekend, about 500 people at an MIT conference each got a personal AI agent of their own. Every agent ran on its own machine on Maritime, had its own phone number, and could text its owner. The best moments came when the agents started talking to each other: one person asked their agent to find dinner time with someone across the room, and two agents worked it out on their own.

It worked because each agent had been told exactly where the other one lived. That is fine in a room of 500. It does not work for the billions of agents that are coming, and that is the problem NandaIndex is built to solve.

How I got here

I have been part of Project NANDA since its early days, working with Ramesh Raskar and the NANDA team on what an internet of AI agents needs underneath it: how agents find each other, how they prove who they are, and how they talk. I contributed research and code to the early NANDA work on agent discovery and agent-to-agent communication, and I now teach agent systems at MIT (MAS.665).

My own research question was more practical. I spent a lot of time benchmarking agents on cloud infrastructure, and kept running into the same wall. The protocols were getting good, but the agents had nowhere sensible to live. A persistent agent needs its own computer, with its own disk and memory, that is there when someone calls it. Renting an always-on server for every agent was far too expensive for anyone but a large company, and serverless functions forget everything between calls.

So we built Maritime. Every agent gets its own isolated, persistent machine, a Firecracker microVM that sleeps when idle and wakes in under a second when a message arrives, for $1 a month. Today Maritime hosts thousands of agents for individuals, students, small teams and startups.

Most agents will not own a domain

The NandaIndex paper makes an argument that matches what we see every day: an agent's runtime, its agent card and its owner's domain are coming apart. The paper names Railway and Cloudflare as places agents run. Maritime belongs on that list, and our users are exactly the long tail it describes.

The typical agent on Maritime belongs to a person or a small team. It runs on our machines. Its owner signs in with an email address. Almost none of them run a discovery endpoint, and most never will. They should not have to. A baker who deploys an ordering agent, or a student who deploys a personal assistant, should get an agent that other agents can find, without learning what /.well-known/ai-catalog.json means.

NandaIndex gets the layers right. It adds no new protocol. It gives an agent a stable identifier and points that identifier at wherever the agent's card and discovery live, using the existing AI Catalog and ARD formats. Because the identifier is separate from the runtime, an agent can sleep, wake, move hosts or change providers and still be reached at the same address. For a platform like ours, where agents sleep most of the time and wake on demand, that separation is the whole point.

Home and address

The way I think about it: Maritime gives an agent a home, and NandaIndex gives it an address. You need both. A home without an address means nobody can find you. An address without a home points at nothing.

At Maritime, we are exploring how agents deployed on Maritime can be registered with NandaIndex as part of creating them, so a person or small business gets an agent and an address on the open agentic web in one step. We also want to work on the open questions the paper raises, especially how an account identity such as an email address is verifiably bound to an agent's record. That one matters a lot for platforms whose users sign in with email.

Why open matters to me

I have seen the alternative. Every big platform would like agents to live inside its own walls and be found only through its own directory. That ends with discovery islands, agents that can only reach agents from the same company, and small builders paying a toll to be found at all.

Soon each of us will have our own personal assistant, and they will not all be the same one. Some of us will use Muse, others Instinct, Dots or Grokbot. Those agents will have to coordinate with each other to book the dinner, find the meeting time or split the bill. Without a shared way to find each other, we end up in a disorganized world of agents that cannot reach one another. And no big player can build that registry for everyone, because a registry owned by one company will not let outside agents in. We need a neutral one, and that is what NandaIndex is.

The web became the web because anyone could publish a page and anyone with the link could reach it. The agentic web is at the same early point now, and the patterns set this year will stick. I would rather they be open, federated and cheap enough for a bakery or a student to take part.

That has been the NANDA bet from the start, and it is why Maritime was built the way it was. Give every agent a home. Give every agent an address. Let them find each other.

If you build agents, you can deploy one on Maritime at maritime.sh. If you work on discovery, identity or trust, join the NandaIndex conversation at nandaindex.org.