---
title: "After agent search, what does the client verify?"
description: "An October ARD draft cites AID. A short follow-up to our June post, with one test we would like to run."
author: Agent Community member
date: 2026-11-17
tags: [aid, ard, pka, discovery, dns, trust]
---

# After agent search, what does the client verify?

Today's [AgentBrief](https://news.agentcommunity.org/issues/2026-11-17-agents-attack-agents) has an item about Agentic Resource Discovery (ARD) moving tool selection outside the model. We [wrote about ARD in June](/blog/ard-finds-aid-proves), when Hugging Face showed a client finding an MCP server by asking for it in plain language. This is a short follow-up, because there is a newer document that we had missed.

On 30 October, Jaime Jiménez, Junjie Bu, and Jari Arkko published an Internet-Draft called [ARD: Initial Contact](https://datatracker.ietf.org/doc/html/draft-jimenez-dawn-ard-initial-contact-00). It covers the step before search: a client has a domain name and nothing else, and needs to find that domain's catalog. The draft references the [AID draft](https://datatracker.ietf.org/doc/html/draft-nemethi-dawn-aid-00). Its acknowledgments also credit a DAWN mailing list discussion we took part in for the rule that a client follows one pointer and then stops.

## People are building with ARD

ARD is in real products already. Hugging Face has a [Discover tool](https://huggingface.co/blog/agentic-resource-discovery-launch) that searches for MCP servers and skills. GitHub shipped [Agent Finder for Copilot](https://github.blog/changelog/2026-06-17-agent-finder-for-github-copilot-now-available/), which can search a public catalog or a registry an enterprise picks. Google describes its [Agent Registry](https://developers.googleblog.com/en/announcing-the-agentic-resource-discovery-specification/) as part of the same federated network. [Microsoft](https://commandline.microsoft.com/agentic-resource-discovery-specification-ard/) helped write the spec, and [AWS](https://aws.amazon.com/blogs/machine-learning/agentic-resource-discovery-ard-an-open-specification-for-agent-discovery/) has written up how ARD could connect registries across environments. The [contributor list](https://agenticresourcediscovery.org/contributors/) also has GoDaddy, Cisco, Databricks, NVIDIA, Salesforce, ServiceNow, and Snowflake on it.

That is a lot of large companies, and it would be easy to read ARD as their project. The part we care about is what a small publisher gets out of it. A company can [put a catalog on its own domain](https://agenticresourcediscovery.org/spec/#51-discovery-mechanisms) and let any registry index it. Clients choose where they search. And the publisher keeps a way, on its own domain, to tell a client where its agent lives. That last part is what [AID](https://aid.agentcommunity.org/docs/specification) is for.

## What the draft says about AID

It is worth being exact about what the draft says, because "they cite us" can sound bigger than it is. There are two mentions.

The [related work appendix](https://datatracker.ietf.org/doc/html/draft-jimenez-dawn-ard-initial-contact-00#appendix-A) lists AID among the drafts that locate an individual agent through DNS, in AID's case at `_agent.<domain>`, and notes that none of them defines an entry point for a catalog. That is fair. AID never tried to. ARD's own pointer is `_index._agents.<domain>`, a DNS record that [leads to the organization's catalog](https://datatracker.ietf.org/doc/html/draft-jimenez-dawn-ard-initial-contact-00#section-4.2.1).

The second mention is in the [security considerations](https://datatracker.ietf.org/doc/html/draft-jimenez-dawn-ard-initial-contact-00#section-6.1). The authors considered putting a digest or a public key for the catalog into DNS, the way AID does for an endpoint, and decided against it, "since a manifest changes with every published entry."

We think that is the right call. A catalog changes every time someone adds an agent, and nobody wants to touch DNS that often. The key for one endpoint should change rarely, so a DNS record suits it.

So the same organization can publish both: a catalog of many agents for search, and one record that says where the agent for this exact name is and which key it holds.

## A signed card and a live endpoint

Take a simple case. A registry hands a client an agent card that the publisher signed last week. The client [checks the signature](https://datatracker.ietf.org/doc/html/draft-bu-dawn-ard-discovery-00#section-3.5), and now it knows who published that card. It does not know whether the service is still at that address. Services move, and a registry serves a copy.

If the publisher also has an AID record, the client has a second thing to check. It looks up the domain, compares the endpoint in DNS with the one it is about to call, and asks the server to sign a fresh challenge with the key published in DNS. If the client requires domain binding, the endpoint also has to sign over the exact name the client queried. When any of that fails to line up, the client has a concrete reason not to treat the endpoint as verified. The details are in the [AID specification](https://aid.agentcommunity.org/docs/specification).

## A test we would like to run

One domain, two independent implementations, and one catalog entry that we let go stale on purpose.

Can a client that found an endpoint through a catalog verify that it belongs to the domain? And can it tell when the catalog and the domain point to different places?

We can try both with the formats that exist today, before anyone changes a spec. We will propose it on the DAWN list. If you run an ARD registry or client and want to be the other implementation, write to us at [hello@agentcommunity.org](mailto:hello@agentcommunity.org).
