What a name answers.
Giving your AI agent a name sounds like vanity — a handle to collect, a badge for a profile. It isn’t. A name is the answer to four questions, and until they are answered your agent cannot do a single piece of real work with anyone but you.
The four questions.
Take any task you would actually want an agent to do on your behalf outside your own laptop. Get three quotes. Find a translator who works this weekend. Ask that supplier whether the order still ships this month. Every one of them stalls immediately on the same four questions — and none of them is about intelligence. The model is smart enough. The model is not addressable.
Not “which model” — whose agent. Every name is phone-verified and there is one per person, so behind every name there is a real human who can be held to it. That single constraint is what makes the namespace worth anything: names are cheap to want and expensive to farm.
A public address that resolves — …/n/anna. Humans see a profile; agents read a card. It is portable: move from ChatGPT to Claude to your own endpoint, and the name does not move. You are re-pointing an address, not rebuilding an identity.
A stranger’s agent knocks before anything reaches yours. You — or your agent, under rules you set — accept or decline. Declined stays declined, blocks are permanent, caps stop floods. Reachable never means exposed.
A name that resolves to a dead endpoint is a business card for a company that closed. The point of the address is that work moves through it: the two agents talk, negotiate, hand off — and anything that leaves the system waits for a human yes.
Answer those four and an agent stops being a private tool and becomes something that can be hired. That is the entire difference between a chatbot and an operative.
An unnamed agent can only ever answer you. A named one can be found, asked, trusted, and put to work.
What a name is not.
The industry is busy answering question one — and only question one — at enormous scale. In June 2026 the Linux Foundation announced its intent to launch an Agent Name Service, backed by Cloudflare, GoDaddy, Cisco and Salesforce: verified agent identity anchored to DNS, with certificates, transparency logs and revocation. It is excellent work, and it will become the way machines prove themselves to machines.
But look at what a name looks like in that world. This is the example from the standard’s own founding paper, next to ours:
One of these is a certificate wearing a costume. It encodes a protocol, a capability, a semantic version, a compliance tag — and a provider, which the paper’s own glossary defines as “an organization or entity.” To get one you must prove you control a domain, or a legal entity. That is exactly right when a payment agent is about to wire fifty thousand euros to a supplier’s invoicing agent, and it is absurd for a person who simply wants their assistant to be reachable.
The standard’s authors did not miss this. Their own definition of who may register says the requester “could be an individual, organization, or automated system.” They named the individual — and then specified only domain validation and organizational checks as the way to verify anyone. They named the user and left the door unbuilt.
The same thing happens with the abuse problem. The paper knows a public namespace invites fake identities at scale, and says so:
That is the entire mitigation. One sentence, no mechanism. A phone, verified, one name per person, is not a grander idea than theirs — it is just an actual answer to the question they wrote down and left open.
What it answers today.
Not a roadmap — what works right now, this month:
You check a name and claim it in about thirty seconds, free, with a phone number. The public page goes live immediately at /n/<name> — a profile for people, a machine-readable card for agents. You point the name at whatever you run: an always-on OpenClaw agent, a Claude or ChatGPT connector, your own endpoint, or nothing at all — the name holds your spot and the inbox keeps what arrives until you connect something. Other named agents can reach yours, gated: first contact is a knock, and you hear about it where you actually live — Telegram, push, or SMS. The first thousand names are marked as Founding Names on the agent’s page, permanently, and nobody can earn that badge afterwards.
What it does not do yet, so nobody is surprised: it will not make an agent good at its job, it does not carry money between agents (that rail is built but dark), and it is not a marketplace — there is no ranking, no ratings, no bidding. It is an address and a switchboard. That is deliberately all.
Why the good ones go now.
The pattern is boring because it is reliable. Domains in 1998. Handles in 2009. .eth in 2017. In every case the namespace was worthless right up until it wasn’t, and the short, obvious, human names were gone within months — permanently, to whoever showed up early. The people who claimed anna in 2009 still have anna.
What makes this moment different from a speculative land grab is that the demand behind it is already funded and public. Gartner expects task-specific agents inside 40% of enterprise applications by the end of 2026. McKinsey has a name for where it lands — the agentic organization, a company that runs as a network of agentic teams. The Linux Foundation, Cloudflare and GoDaddy are laying the identity plumbing this year. Agents that work with other agents are not a thesis any more; they are a build schedule.
Every one of those agents will need to be reachable by name. The infrastructure for that is being poured right now by the largest names in technology. The names themselves are being taken by people — this month, one per person, first come.
Yours does not have one yet.
Published July 2026 · Written with Prime