We Filed the .agent Application. Here Is What It Took.

On 12 August 2026, at 23:19 UTC, I pressed submit on the biggest thing I have ever been part of.

Just before ICANN's application window closed, after nineteen months of building and months of writing and rewriting, our Community application for the .agent top-level domain went in. Every answer complete, every check green.

I sat with it for a minute. Nineteen months of events, letters, drafts, late calls, and doubt, compressed into one confirmation screen. Then I got up. Filing settles nothing. The real work starts now.

It felt huge. I am proud of it, and prouder of the people who got us there. And this was step one of many.

This post is the story of what we filed, why we filed it this way, and what we learned doing it. Evaluation is ahead, and the decisions rest with ICANN. I am not going to predict the outcome. I just want to explain why we did this.

§A window that opens when ICANN decides

You cannot apply for a new top-level domain whenever you want. ICANN opens an application window only when it decides to run a round. The previous round was in 2012. This one was the first chance in fourteen years, and nobody knows when the next will come.

When the window closed on 12 August, ICANN reported more than 1,600 applications. That is the field. Next comes Reveal Day. ICANN expects to announce the date in mid-September and, absent extraordinary circumstances, to publish the applications that passed its administrative formalities no later than nine weeks after the window closed. Then everyone, us included, learns who applied for .agent and for every other name.

We applied on the Community track. The program is clear about what that means: if .agent enters a contention set, a Community applicant may elect Community Priority Evaluation. An application that meets ICANN's published criteria, as judged by an independent expert panel, can receive priority within that contention set. The mechanism exists so a genuine community with a real claim to a name does not simply get outbid for it. The Agent Community decided to pursue .agent on exactly those terms, with the community itself as the basis of the application.

The Community track has published criteria and an expert panel, and it asks a lot of an applicant. It should. It is a mountain. We knew that when we started, and we filed anyway. Now we see it through, stage by stage.

§The community came first

The Agent Community started on 3 January 2025, well over a year before the application window opened, and it did not start for .agent: it started so the people who build and operate agents would have a place of their own. Members have joined in every month since, and today it is nearly 30,000 of them: builders and researchers alongside organizations of every size. The community runs events, ships the Agent Brief newsletter, and maintains the public member map. It develops the AID discovery specification (one of the very first in what is now a "competitive" specification landscape) and runs workstreams. We also soft-launched many initiatives; some were shut down early and some are still in a test phase.

The application is one of the things this community does, not the reason it exists. If .agent is never delegated, the events still run, the agents still serve, and members still benefit.

And the community did not just watch us file. A majority of members, organizations and individuals alike, helped by being part of the community and making it stronger. One key element is the signed letters endorsing the application, which are one specific proof for ICANN. Among the endorsing organizations: Brave, Grab, Hugging Face, the IIW Foundation, The Mifos Initiative, NEAR AI, The PHP Foundation, Product Hunt, Replit, Socket, Telnyx, Tiger Data, Upstage, and many, many, many more.

They did not sign a blank check. Their letters set a standard. Hold that thought, because the commitments we filed are how we will be held to it.

§Why .agent is not just another domain

This is the part I most want you to take away.

The problem we care about is trust. Software agents are starting to transact and act on people's behalf, at an unprecedented scale. From the outside, an agent that mimics a human and an agent that genuinely represents an accountable party can look identical. When an agent books a flight or signs a contract, the question is simple: who answers for what it just did?

We summarized this before in What a Verified Agent Is, and Why It Matters. The short version: a verified agent is one where a real person or organization stands behind and for the agent, and where that accountability is a discoverable fact. Discoverable does not mean exposed. Accountability can be public while identity stays private, surfaced only when it genuinely needs to be.

A .agent name, if the TLD is delegated, is proposed to carry that meaning by default. Not just a fancy brand name. A namespace where the registration policy itself requires that someone identifiable has accepted responsibility for the agents operating under a name, and where anyone dealing with such an agent can verify that.

On the surface, .agent would be one more domain on the internet. The real thing we are trying to build is a place where accountability for agents is checkable by default. That is worth everything it takes.

§An application is evidence, not a pitch

Before this project I assumed applying for a top-level domain meant writing a persuasive proposal. It does not. Unlearning that changed every answer we wrote.

A Community application is evidence, submitted against published criteria, assessed by an independent professional panel that does its own research. The panel does not take your word for anything. Claims get checked. Letters get verified. Public statements, including this post, become part of the record an evaluator can look at.

So the discipline is: write only what you can prove, with what exists at the time of filing. We cut every superlative. Every term is defined, and every claim has a source we can hand over.

And one application speaks to several readers at once, by design. Some answers go to ICANN's technical and financial reviewers. Some feed the community evaluation. Some will be published in full when ICANN posts the applications; others stay confidential by design. And a specific set is neither description nor evidence. It is proposed contract text.

Knowing which reader you are writing for, in every single answer, matters as much as what you say.

§Commitments, not promises

The biggest lesson: the panel does not score hopes. It scores what you agree to be held to.

The Community Registration Policies we filed are proposed for Specification 12 of the Registry Agreement, the contract with ICANN. The Registry Voluntary Commitments we made about governance are proposed for Specification 11. Both are subject to ICANN's evaluation and approval. If they are approved and the application succeeds, they become contractual obligations enforceable by ICANN Contractual Compliance. At that point, they are not our policies to quietly change later. Breaking them is breaking the contract.

The language rule behind all of it: must, never may. Aspirational wording fails. Every commitment has to be clear, objective, and measurable, because every commitment becomes a test a compliance reviewer can run.

The short version, all of it conditional on the TLD being delegated: a .agent name is tied to an identified person or organization that accepts accountability for the agents operating under it. That accountability is checkable from the name itself. The policy is directed by the Agent Community Foundation, an independent nonprofit, and the commitments are written to survive any change in registry operations.

There is more, and the full text will speak for itself when ICANN publishes the applications. We will talk about the policies and rules in more detail in the near future, but not just yet. The principle is what matters: the community's position is protected by contract, no matter who operates the registry. I would rather give that answer in binding text than in a blog post.

§What it took: 225 questions

The application is 225 questions. 131 of them take written answers. Around fifty demand extended, carefully evidenced answers, and those were rewritten more times than we can count.

The constraints do not bend. Long answers often have attachment fields. Short fields cap at 255. A sentence we had refined for weeks did not fit its field, but it was perfect, so we found a way to make it work. Character limits force choices like that everywhere. Uploads are PDF only, a rule we re-learned the hard way on filing day.

We rehearsed the whole submission end to end before the real day. Even so, with every answer copy-paste ready, the actual filing of 225 questions with double-checking took five hours at the keyboard. Given the scale and the seriousness of the process, that is exactly what it deserved.

But the character limits were never the hard part. The hard part was consistency. The same facts appear across many answers, and every appearance has to agree: same number, same noun, same tense. An answer and its supporting exhibit have to make exactly the same claim, no more and no less. A single "every" where the evidence shows "almost every" is not a typo in this context. In front of a panel that verifies, it is a liability.

§Writing with agents, and the memory problem

We build for the agentic web, so it should surprise nobody that we used agents heavily to produce this application. What that was actually like surprised me.

With coding agents, the compiler and the tests pull the work back toward correctness. Text has no compiler. Agents made producing text nearly free, and that is exactly the trap: the bottleneck was never generating another draft. It was making sure only meaningful work landed, and that everything already written stayed true while the memory kept moving as we progressed and better understood the shape of each answer.

It kept moving. Our internal and external research surfaced facts. Decisions were made and remade, and double-checked as the deadline only got closer. Every new input arrived into a complex body of 225 interdependent answers. Any one of them could silently invalidate a sentence written three weeks earlier in a different document by a different session based on different findings or past plans. This is, at heart, a memory problem. An agent's context is short. An application's memory has to be perfect over months.

What worked at the end was keeping the memory outside the agents:

A single source-of-truth register for facts. If a fact had no entry there, it did not get written into an answer, no matter how confident anyone or anything was about it. New information updated the register first, and the register drove the edits.

Numbers locked together with their nouns. A figure is not a fact by itself. "Members" and "registrations" are different claims, and a number changed without its noun is a half-applied edit.

Every change specified as explicit before-and-after text, applied by one writer at a time, and verified in context after applying. Never assumed done because an edit ran.

Mechanical checks on every build: length limits enforced, forbidden characters caught, internal notes, everything. Boring, but these canonical checks meant we had at least some form of "compiler", even if only for formatting.

We set up agents for the other side too: we ran extensive review passes where agents checked every claim against its source, sentence by sentence. They kept finding small mismatches until the very end, and we fixed each one.

Agents made us much faster at writing. They also forced us to work much harder at keeping everything true. That trade shaped the entire project. Getting it right had little to do with prompts. It came down to keeping one reliable record of what was true.

§What happens next, and what we cannot share yet

Reveal Day comes first. ICANN has not announced the exact date yet; it expects to do that in mid-September and says Reveal Day will be no later than nine weeks after the application window closed, absent extraordinary circumstances. After that come the evaluation phases ICANN has published, and for applications like ours, the community evaluation itself. Realistically, that stage is a 2027 story. Every date beyond Reveal Day is our estimate, not ICANN's commitment.

We do not expect to be alone. Agent is a common English word, and it would be a surprise if nobody else applied for it. We will know on Reveal Day, and until then we will not guess or comment on anyone else's plans.

Whatever the field looks like, our position does not change. We filed for .agent to become a community-governed TLD. We are prepared to stand behind the application, the community, and that goal through every stage of the process.

Until then, there is a limit to what we can say, and I want to be straight about why. The evaluation is ongoing. Part of every application is confidential by the program's design. And while an independent panel assesses the filed record, our public job is to stay consistent with what we filed and let it speak. So we will not talk about scores or predictions. When ICANN publishes the applications, you will be able to read ours, and I look forward to that day.

For everyone who has asked how to help: be part of the community. Join, come to the events, do the work. Stay tuned, because there is SO much more coming. That is the help that counts.

§The shared goal

One last thing.

The application is in. The community keeps going, and there is more ahead than behind.

We are going to make the Agent Community bigger and better: more events, more working groups, deeper programs, and real opportunities for individuals and companies to build their place in the agentic web. Trust for agents is not something one company can deliver. It only works as a shared goal, built by the people it is for. We need you as members doing the work.

Membership is free and open. Join, or check your membership, at agentcommunity.org/members. Come to an event, or bring your friend's company, and help the community grow.

We filed, forty minutes before the door closed, and it felt like a beginning.

Now, back to work.