Google Agentic Resource Discovery (ARD) Explained: The new AI Agent Standard

[00:00:00]So, let's talk about this brand new thing called agentic resource discovery.

[00:00:04]It just dropped this week and it's not a one company thing. Google, Hugging Face, [music] and a bunch of other big names put it out together. So, here's the weird thing it actually goes after, right? So, an agent can only use a capability it already knows that exist.

[00:00:18]When I say capability, think about tools and memory and all of those interesting things, right? But, for the agent to actually use it, it should know that exists, right? Even if it has not heard of it or if you have not provided that information, it will not be able to use it. Now, they've built a lot of pieces for agents lately, right? You obviously know about MCP, which is how agent plugs into your tools and data. Then, you obviously know about skills, which is packaged like know-how on what the agent should do so that it can keep picking it up. And then, agent-to-agent protocol, which is how one agent can talk to another agent, right? Like, all of these things have already been built, right?

[00:00:55]So, tools, skills, and agents themselves. Now, if you think about this, these are all the same ideas broadly, right? So, these are capabilities your agent can reach for.

[00:01:07]There's a name for the whole bucket, which is called agentic resources.

[00:01:11]And every one of those pieces is about how an agent uses a capability.

[00:01:17]None of them answer how it finds the one out of thousands of capabilities that exist, right? And obviously, you cannot cram 10,000 tool manuals into the agent's head and hope that it just works. It just does not fit. So, the hard part actually now stopped being how an agent uses a tool, it's actually like how the agent can find the tool. And that's what ARD is going after. So, your agent asks one question, "What's out there that can help me do this?" and gets back real answers. What each thing does, who made it, where it lives, and how to reach it, right? Another thing that this is very serious is one of the people who actually wrote this, created RSS and schema.org. So, the standards that taught the search engines how to read the whole web, now that individual is actually helping the same thing and creating same thing for agents, right?

[00:02:08]So, the best way is always to get to understand this is to watch this visually. So, what I did was I built this demo where I built actually two parts. First, an agent tries to find and will try to ask a question of the agent, but it will not be able to find it and then I will use ARD in the second part and the agent will be able to discover it, right? So, a quick disclaimer before [music] we get into it, all opinions are my own and do not belong to my employer.

[00:02:33]All right, with that, let's get [music] into it. So, before we get into the demo, I quickly wanted to take you through the blog post and a really good architecture diagram that I found here, right? You can see the announcement here and you can see all the different partners who've actually helped to build this, Google being one and then there are other prominent companies also listed here.

[00:02:52]But and then rest of the things is what we went through earlier in the video.

[00:02:57]But this is a diagram that I really liked and I think it will help us understand this concept really better.

[00:03:03]So, if you think about this, there are two different parts to this. One is the catalog and the other one is registry.

[00:03:10]So, think about the catalog as any company that is describing its agent or tools that could be skills or MCB servers in a single file and that file will be the AI-Catalog.json, right? And then they will host it on their own domain and there's no registration required. So, think about this as the crawlers will be able to crawl. So, in this case, the agent will be able to crawl and read this, right?

[00:03:34]So, that's what really what a catalog is and it lives on the domain, right? Now, on the right-hand side, we have the registry. So, registry is a very important aspect because this is where you are providing a discovery service that crawls lots of these catalogs and indexes them for you, right? So, your agent is going to ask by intent. So, what can help me do this particular task, right? And then it will get back basically ranked matches from the registry, right? So, really two big reasons this exists. An agent cannot hold all of those thousands of tools description in their head, right? So, the search happens outside the model and a registry can layer on richer signals like ranking, trust, policy, and all of those things, what's approved. And then agents can register and refer you to other registries as well. That's a cool thing that we you also need to understand that, right? So, that's the idea and then both land on your AI agent and that's where you can see that both the arrows point over here, right? The agent will filter the ranked matches, will verify the publisher, you can see all of these things are also listed here, and then authenticate, and then invoke. The agent will invoke actually like which ones will be the protocol or that it needs to follow, whether it's an e2e or whether it's an MCP or whether it's an as simple as like an API call, right? So, you can see that broadly there are five steps here. So, you as a company you publish your catalog and then your registry may index the catalog and then the agent will basically search the catalog or the registry and your agent will verify and then basically invoke and that's the whole connection piece, right? So, that is really what this diagram is talking about and in in a nutshell what ARD is all about. Now, this will become even more clear when I actually take you through this, right? So, this is a demo that I just quickly built using anti-gravity and I'll also share the link to the repo that that I actually used, which which actually helped me a lot to to build this one. So, this is the ARD discovery demo and then you have all of these specifications. So, if I open my anti-gravity for a second here, you can see that all of these So, this is like a real agent that I've created and you can see the ARD spec and everything, which is mentioned over here, right? So, this is as real as it could get. So, the idea here is if I come here then I can actually ask that hey, what is the weather in Chicago?

[00:05:55]Okay? So, I'm asking this question because as of now this particular agent only has these two tools, right? Calculator and echo. It doesn't have a tool for it to access weather from Chicago, right? And that's where the tool could be from let's say weather.io and weather.io would go ahead and register as I was explaining their catalog and then make it available in the registry. So, now let me ask the same question here, right? So, now in this one I have actually enabled ARD. So, let me ask this question, what's the weather in Chicago?

[00:06:32]So, now what it is doing is it is going to look for you know, basically doing the discovery, right? So, here are three manifests that I already I've shown here. So, this is acme.com's manifest and then I have the weather.io's manifest as well, right? So, there you go, AI catalog.json and then advanced martech is mine, right? That's my website. So, I just wanted to show like it could get as real as this. So, here it would find that okay, it is going to do the crawling. So, all of this it is happening as we had seen, right? is all the steps that is hap- happening here.

[00:07:03]So, now the agent is actually querying, basically sharing the intent and based on which is the right particular tool or resource it has already identified and now that it has identified then it hands it off to A2A, right? And that's where the A2A will actually go ahead and write the result and you will get an answer that this is the weather in Chicago, right? So, this is really like the end-to-end demo here, right? I'm just going to run this again really quickly so that you can see the end-to-end flow, right? What is going to happen is the manifests published from different websites. Once that is done, then it will crawl and index. That is an optional step and then it will all be there here, and that's where the agent sends the query. Then, it'll verify the identity and also the result, and then it will hand it off to E2A. So, pretty cool. And this will actually really help when you have thousands and thousands of pages, which is going to happen if not now, but very soon, given where we are in the world. Another thing that I saw was, you can see that there is an agent registry here, which I think the Google Cloud team put together. So, let's say that you're a large company and you're creating all of these different types of agents and multiple agents of for your own company, instead of pushing 10,000 different tools in the agent's head, you can actually maintain a registry like this, right? So, this is pretty cool.

[00:08:22]And this is not the only one, right?

[00:08:24]Like, we have other companies who have also done this. GitHub has its own, Hugging Face has its own, but I just wanted to show you a a quick example of what I was able to find, right? Which is obviously, I was able to find this as part of the blog here. So, that's what I wanted to explain today. I think as we are moving forward with so many agents working together and this whole trying to really push a lot of context, a lot of tools, a lot of skills to individual agents, this kind of ARD or agent repository discovery, agentic repository discovery is going to be super super valuable. So, I would suggest that let's pay close attention to something like this. It's always good to have this in our knowledge arsenal because these are all upcoming technologies. So, hoping that this video added some value to you.

[00:09:10]So, as always, let me know in the comment section if you guys have any questions, and of course, if you have any other comments as well. And if you're new here, please do subscribe to the channel and give me a thumbs-up so that I know that this is resonating well. Thank you very much for your time.

[00:09:23]Thank you for watching, and I will see you in the next one.