Google Cloud Next 25: Building and Managing AI Agents A2A Protocol

E65 GOOGLE CLOUD NEXT 25_ BUILDING AND MANAGING AI AGENTS A2A PROTOCOL

Transcript

Dalton Anderson (00:01.71) Welcome to Midget Step Podcast, where we discuss entrepreneurship, entry trends, and the occasional book review. Google just had their Google Cloud Next 2025 conference in Las Vegas, and I watched the two-hour, very dense video and demos, and something had caught my eye. There's a couple things, but for this episode, we'll be discussing one. And I think it's one of the...

last pieces for agents to actually scale. And it is the agent to agent protocol. And what is that and why is it important? And what does that mean? It's everything that we'll be discussing in today's episode. Once again, I'm Dalton Erson, your host, of Interstep Podcasts. So Google had a wonderful

conference and a keynote that I think was around an hour and 40 minutes. And a lot of the information was industry leading or

incredible the stuff that they were announcing. And it really truly shows how much money they've spent on AI and their technology stack being vertically integrated from servers to building their own chips to infrastructure, to oceanic fiber networks, to the cloud infrastructure that they're able to provide to themselves and to their clients, to the video and

text data that they're able to train on, that being YouTube or Chrome.

Dalton Anderson (01:53.686) all of that stuff combined into a wonderful AI product. And then now that AI product is industry leading, Gemini 2.5 Pro.

as of like 10 days ago was number one on chatbot arena. Their models are for compute costs and performance are the best performing models on the market. And so Google is just killing it in so many different areas. And one thing that was important was the agent to agent protocol.

And I'll get into why that's important in just one moment. I just wanted to give a shout out to Google for their wonderful keynote. And another thing, this is sidebar, side personal bar conversation that I had at a meeting this week was I joined the meeting at work and I say, they're like, hey, darling. And then I'm like, hey, everyone. And then they're like, whoa, whoa.

Look at you with your podcast voice. It sounds like you have a podcast. And I was like, yeah.

Dalton Anderson (03:13.932) Yeah, I guess I have a good voice for that. And then I kind of just kept it low key, didn't say a thing about it. And then somebody in the meeting piped up and was like, well, actually, Dalton has a podcast. And they're like, really? He does. And I was like, yeah, I do. And then.

Eventually everyone found out my podcast and they said they followed and subscribed either or depending on the platform. So we'll see if they're listening in on this episode. They could be here in the real or not the real world. I was going to say in the nether world, but anyways, they could be here listening out here somewhere in the world through the internet. So if you're listening, appreciate your listen.

and hopefully you find this content interesting. Thanks for following and supporting. Okay, so now we're going to get into agents, agents protocol. And I have my little notes here. Just want to make sure that I am all good. So before I talk about the agents agent protocol, there's another thing that's also important to talk about and that'd be Anthropics model context protocol and Anthropics MCP.

is an open protocol that allows models to

integrate with external tools, either be Google Maps, the internet, and that could be Microsoft or Chrome. But basically, instead of building in custom code to embed your product into another product, like your AI model, if you want to look at financial data, you don't have to build out each company doesn't have to build out.

Dalton Anderson (05:15.406) custom code to embed their product into S &P Global. What this MCP protocol does is, here are the key places where AI models want to go, and let's just make a single protocol for these AI models to access this information. One, the vendor will be able to potentially be compensated if they want to and whatever matters they want, either maybe free.

AI credits or they're getting monetary compensated. mean, there's a whole bunch of different ways you can be compensated for these kinds of things. Then the next thing is it allows people to easily build on top of what has pre-existed. And that's important because if everybody's utilizing their resources to build the same thing as other folks, then those resources aren't allocated in the most efficient way. And so this MCP allows

models and companies just to use the protocol instead of building out their own say protocol, proprietary protocol for tool integration, external tool integration. It's otherwise known as tool usage. It's a very important piece of providing additional information to the query user. So the person in the chat or the person asking things via API.

And another thing it's really important for is grounding, grounding your answers into truth. So if you're saying that something is open and it how do you know it's actually open? You might have data on it historically from all your training data because you trained on the whole Internet, but things change and that may not necessarily be true now. So previously your dominoes nearby

was open till 10 and that might be true or might not be true. The way you can ground your truth in that statement is verifying it on Google maps or verifying it on Apple maps. And the way to do that is with this NCP protocol, the model context protocol built by Anthropic. And also like to note that Google is an Anthropic investor and also Anthropic trains their data on Google's TPUs or to tensor processing units.

Dalton Anderson (07:39.584) And those are Google's proprietary.

ships that they built for AI workflows. Okay, with all that out of the way, you understand the lay of the land. So there's this tool usage protocol built by Anthropic. And then now there's this other protocol, agent to agent protocol. And so what that allows people to do is it allows agents to communicate with other folks. So agents can communicate with other agents.

And that's a big issue that we had was agents maybe go to like one thing. So you might have a customer service agent and then you might have a, I don't know, shipping processes in the agent and they might not, they might have great abilities in their silo. But once you try to have a complex workflow of multiple agents, if they can't communicate with each other and share data in a safe manner,

and or share data at all, then it doesn't necessarily work. But now what this agent to agent protocol allows is agents can communicate with different systems. Doesn't have to be on your system, which is great. So it's not owned by anybody. It's not owned by Google. Google proposed a protocol, but it's open to the whole world to use the protocol. And Google has a pretty good head start as they have.

a lot of high profile partners for the agent-agent protocol and the most notable one being Salesforce. And they have a really cool demo that I'll share later on in the episode pretty soon. But it basically, breaks down the communication barriers between agents. it allows for more complex workflows and it allows agents to be a feasible technology that can scale.

Dalton Anderson (09:43.382) And then as I mentioned before, this agent to agent protocol is not specified for one framework or technology stack. So if you had your agent built in Google, like you use Google's Gemini and you built it in Vertex AI, and then you had some kind of database stuff in MongoDB or AWS, other, I would say,

other processes or other vendors when they're building these kind of full suite things, it's got to be within their walled garden. Like think about it, like Apple does that great. Like Apple's perfect to that where your iPhone, you only can have an iPhone. And then when you're Mac and you want to communicate between iPhone and Mac, that works great. But as soon as you work outside of the ecosystem, everything breaks. Whereas the agent to agent protocol is built to

work with any technology stack. So you could work with, if we're keeping this technology of phones, your, your Mac could communicate easily with your Google and your iPhone and vice versa. And so everybody communicates with everybody. And not only is that the availability, but then as I mentioned before, is it doesn't force people to consistently utilize their resources on things that other folks will be building.

And so it's like kind of like a build once and that's the way it is. I mean, there probably be some iterations and improvements, but there's less resources spent on building the same thing and people can spend their resources on things that are more important to add more value. So before I get into why is agent agents like important, I want to share my screen. And if you're listening into the podcast, that's great. It's a,

audio and it's a video. So if you're watching or listening, I don't think that will affect your experience. So we do this. So this is Patrick Marlowe. He's the product manager of Applied AI at Google. So I'm to play this video and it's at, and if you want to watch it yourself or you want to skip this and watch it offline or when you get home within the video, Google cloud next 25 opening keynote.

Dalton Anderson (12:12.312) the new way to cloud. It's on the timestamp of one hour and six minutes and 29 seconds. OK. So this video or this demo is going to be a couple of minutes long. But I think it really emphasizes the power of the agent to agent protocol. And this is what I think would be.

the video that you could extract the most value. And if you wanted to leave this podcast episode, I think you'd have a pretty good idea of how it works. All right. So I will share my screen and let's make this full screen. All right.

playing now.

Dalton Anderson (18:07.096) Okay, that is insane. And that's not necessarily two agents communicating with each other, but like I don't necessarily know if those two, there's agents in the backend, I think, where there's a shopping agent and then there's a customer service agent and then there's a connection to your Salesforce instance where certain things that they don't have authority to do, their requests.

approval to their sales manager and that's a human. There's a human intervention or human in the loop and certain transactions or certain requests that they don't have authorization to execute on. And the cool thing about all of that was it seemed pretty seamless. And if you were a manager and you were just getting all these requests, you could respond.

with what you're comfortable with and then the agent would handle it on your behalf. And then from the customer's perspective, everything is happening in async and there's never like a real, seems like a lapse in engagement from the customer's perspective. And I wouldn't really care if I was talking to customer service agent and it was AI and they're able to suggest me the right items that I need for.

And I could share my camera and they can identify which plan I have. Then it tells me I have the wrong soil and then ask me if it's OK if I. Like here's the soil suggestions, is it OK if I switch out the ones you have, you have the wrong ones? That's pretty cool. And then it's also able to schedule schedule an appointment with the the right date.

and time and then also escalate things to the manager and get approvals for things that you would have to wait for anyways.

Dalton Anderson (20:09.16) All of that?

Chef's Kiss, such a cool technology. That at scale would really pay dividends for customer service and allow businesses to focus on the most important things and escalate the things that they need to, but also take care of quite a few niche questions. And I would say.

businesses that deal with niche hobbies like biking, gardening, don't know, parasailing where like things are detailed, but there's not a big community. I mean, I'm not saying there's not a big community in biking, but like it's so complex to figure out which saddle height you need because each. All right, keep this in mind. I'm talking about biking here. I'm switching, kind of switching topics, but to figure out which saddle height you need.

each bike brand and the spacing between the tires. Everything is not standardized. Like everybody's just doing whatever they want. So your height and your, I don't know what it's called, but your, your, your height from like your knee to like your pelvic area, like depending on the length of your thigh bone affects like which height you need plus your actual height. So all of those things affect the sizing of your bike and what would work best for you.

And the issue is each brand has different sizing. And to figure that all out is pretty difficult. And typically, you would get help. Like somebody would help you, like some kind of bike professional, and they either do it for free or they might give you a fee or whatever it may be. But you can ask the customer service agent and they'd be like, OK, like your height is this, your your thigh bone length is this. Then these are the bikes I recommend.

Dalton Anderson (22:06.542) And here they are. you want me to schedule a schedule appointment to try it out, try the bike out and take measurements.

it just, it just frees up so much time and scales in a way where you don't have to worry about people like training people to be comfortable on the phone or training people to reflect your brand in the, in the right way. You don't have to think about emotional instability of your employees or things happening to them that affects how they communicate your brand on the phone. All of those things.

Those headaches are kind of just gone in this scenario if it works the way that it's demoed. And that was a live demo that was 100 % live, wasn't recorded. They did it live in front of the audience.

And I think they're comfortable doing that because they already have 50 partners that are utilizing these AI agents every day and this agents agent protocol. And they've published with their partners, 601 AI use case examples by industry and the company that was using them. So that's why they probably felt comfortable. It's like, okay, like we already, we've already done this a lot.

we've got thousands of agents out there for people to utilize. So that's why I think it's important because the agents in themselves are great ideas, but they're only good if they could communicate with other things. If they can communicate with other agents or other workflows, you'll never be able to build an automated agent workflow without agents being able to communicate with other agents. Like I'm not going to give

Dalton Anderson (23:57.208) you're not going to be able to train an agent to do every single thing, but you could train an agent to do certain tasks very well, but you're not going to have an agent to do everything because then it's not an agent. Then it's just like a human. And maybe we get to that point where it's like an employee, but right now we're in the world of agents and maybe we get to the world of

like executive employee or executive agent or something like that, where that agent has authority over the other agents and schedules work. I don't know. But currently, agents are only capable of being good at like certain tasks when you train them. And so you'll need to communicate with other agents if you want to create a workflow, which is the bread and butter. Like that's what scales. That's the technology that that would have that would allow rapid

adoption and advancement in a short amount of time. And that's what Google is trying to do. And Google is the first hyperscaler to come out with this agent to agent protocol or this open protocol. If you're not familiar, a hyperscaler would be like a Google or AWS from Amazon or Microsoft cloud services, like allows you to scale pretty high, pretty fast. Like if your service requests,

or resources need to scale fast, like that's how you would scale.

So I think that emphasizes why it's so important and the role that Google is playing. they're being an industry leader in this agent technology. They've integrated this agent to agent protocol into their cloud services. And they also have 50 plus partners, like large partners that are utilizing this agent agent protocol that demoed it for some time before it was publicly released.

Dalton Anderson (25:59.884) And so the next thing is now that you have the agents, agent protocol, you've got to make it easy to build agents. And that's exactly what Google has done. within their vertex AI platform, they built a

Agent developer kit or developing kit agent developer kit a decay. They built an agent developer kit. And basically that would be multiple or many.

Many complex agents or agents, I was gonna say complex agent workflows and agents templates that are already pre-built and integrated with key partners. Like if you wanted to build a Salesforce agent or if you wanted to integrate with Workday or these other partners that they called out within their keynote.

That's available right now. And if you're a workspace user or sign up with Vertex AI, you can go in there and you could create your agent workflows. And so then there's that. So they made it very easy to create agents. And the next thing that you need to be able to do is you need to...

Dalton Anderson (27:18.338) you need to be able to monitor the agents and.

Sorry, I'm having some burping going on in mid-episode, so I have to stop what I'm doing.

Dalton Anderson (27:33.804) They built this agent garden where the garden within this garden is where you kind of access all your agents and manage them. So you have the agent developer kit and that's the agent templates or you can make your own. And then you throw it over once it's built or accepted or created. Then you would put it in the agent garden and that's where your agents live. And the agent garden will

Dalton Anderson (28:05.998) will allow you to learn from the working examples they already have inside of the agent garden and also manage your agents. What they showed on the agent dashboard or the agent engine was

A agent request to hire a software developer and that software developer job, they gave it a job description. Then the agent went and found candidates and then the candidates set up interviews. And then afterwards that turned into them accepting the job. I'm sure it'd be more complicated if it was like an actual like situation, but I'm not not ready or not. So I'm not ready. I don't think it's.

something that is as realistic as the demo that was shown where that demo, I'm like, wow, that's blowing my mind there. That's pretty sick. Whereas this hiring of a software agent, or sorry, the agent hiring a software engineer isn't the biggest thing for me.

Dalton Anderson (29:17.088) And sorry, yeah, I misspoke. I'm referencing my notes again. And the agent engine is what allows you to monitor your agents.

Dalton Anderson (29:27.286) So this is the agent workflow or the future agent workflow. What's missing? In my opinion, once agents can communicate with other agents, the next thing is agents being able to conduct economic actions or transactions. And so that would mean that there need to be some kind of agent to agent or agent to worker protocol that allows agents to

hire and transact with other agents or with humans.

That's the last thing. And then from there, you should be able to scale this whole agent economy in a pretty cool way. I think it'd be very cool that agents could hire people or agents could hire other agents. And what manners would I think agents would hire people would be maybe for a gig work or for specialized things that the agent is in Nestle.

an expert in and they didn't feel comfortable with the other agents on the market or something like that. But I can definitely see agents hiring humans to do. At the moment, at least like manual labor, because the agents are virtual, they don't have interaction with the physical world. Maybe they can rent a robot and then the agents could go and do the work like.

Maybe there's these robot, like there's a company that just builds out robots all across the country and then agents could pay to rent the robot to do tasks that that they're asked to do. So like maybe your agent is a shipping agent and so they're at like some kind of warehouse where your products are and there's robots that are like at this warehouse kind of complex. And then they just V.M. into this robot. The robot gets turned online. It does its thing and then it goes back.

Dalton Anderson (31:28.722) or it could pick things up and ship them out or whatever it may be. I think that'd be pretty interesting. Who knows? You might be thinking from what I'm saying is like, wow, that's crazy. That's so far out. But I don't think it's as far out as you think. I think it's closer than it is further away. All of these sci-fi like interactions that I'm talking about. I think you'd truly be surprised.

how close these things are versus how far they were away, how far away they were two years ago. think, I think three years ago, would you say that agents could have a full on phone conversation and it not be, hello, my name is, I don't even know how to do a robot voice. So I'm going to try. I tried. It was horrible. I was like,

Beep beep beep beep boop boop. Yeah, this is I'm embarrassing myself live. Agents can have communications without having a synthetic robotic voice. Agents can request your access to your camera and look at what you're showing them and them understanding and comprehending in real time. Agents then understanding that they have the wrong soil.

in their shopping cart and then suggesting that they change their shopping cart from the image that that was shown. Would provide all the information regarding the shopping cart, request a change, add the changes. Once they're accepted, then the user requests for installation and if they have price matches and then it realizes it doesn't have enough authorization, so then it makes a real-time request to its service manager or account manager.

And then that account manager accepts the request and just reduces the discount. And then the agent knows that, okay, I can't do 50%, but I could do 20%. Here's the updated offer. And then it schedules it for them. All of those things three years ago.

Dalton Anderson (33:48.878) If I would have told you three years ago, three years from now, this is what's going to be live demoed. Like we'll have that capability.

People, people would not believe you. People would not believe you. If like a lot of the things that we have now, these technologies that are being built and are rolled out and they're massively available, it's just people aren't utilizing them. People would not believe that you had that access three years ago or five years ago. People would say you're insane. That you're overly optimistic about what is deemed to be one of the most complex problems in the world.

It seems that it seems to people are keep solving it. Like, I don't know, every time I turn around, there's some kind of crazy AI news and or crazy news regarding technology and advancements. mean, companies are spending $100 billion 75. think Google is spending 75 billion capex every year on AI. And so there's gotta be there's gotta be progress.

a lot of progress or the budget needs to get slashed. And it seems like they're just making a lot of progress. 75 billion is a lot of money. And that's just Google and Nvidia is pumping in crazy amount of money. Meta is doing the same thing. OpenAI is doing the same thing. Anthropic. mean, the list goes on and on and on. And so there's hundreds and hundreds of billions of dollars every year getting pumped into AI. And we are getting, it seems like

at a minimum every three months. There's a really cool announcement. And it was quiet for a bit, but now it's just coming really fast.

Dalton Anderson (35:37.864) XAI acquired X, which seems like a pretty odd deal, but kind of the same thing in the first place. But they're already integrated and partnered with each other. then XAI has also announced their stuff that is pretty advanced. I think recently, as of yesterday when I'm recording this, they had announced

you can now share your camera live with your GROK app or with X, think, and it will allow you, then, let me, I have to check on that.

Just your takeaway from that is you can share your camera. I don't know which app. Well, you can share your camera and Grok can understand what's going on in your environment, real time.

I mean, just the level of advancement that XAI is having within such a short amount of time. Like OpenAI has been working on their AI project for a long time. Google has been doing it for 10 plus years. XAI has been around for like three years. That's it. And they've already caught up. And in some places, they have passed their incumbents.

So just want to keep you in mind or keep that perspective in mind when you're thinking about what's next and what's the next crazy thing. I don't know. I'm surprised every day, the stuff I'm reading and the stuff that I'm talking about. So I just want to make sure that it stays exciting and we just keep pushing for a better, better world. But of course, wherever you are in this world, good afternoon, good morning, good evening and

Dalton Anderson (37:30.734) Thank you for listening and I can't wait for you to listen in next week. Have a great day. Goodbye.