GOOGLE CLOUD NEXT: SCALING AUTONOMOUS AI WORKFLOWS

E113 GOOGLE CLOUD NEXT_ SCALING AUTONOMOUS AI WORKFLOWS

Transcript

Dalton Anderson (00:01.422) Welcome to Venture Step Podcasts, where we discuss entrepreneurship, industry trends, and the occasional book review. I'm your host, Dalton Anderson. And in this episode, we're going to be discussing Google's recent event, Google Cloud Next. And then during this event, they had a couple of things that really caught my eye and my ear. And one of those was of new code created in 2026, 75 % of that is AI generated, which really emphasizes where we're going.

And we're going into this agentic error. And this is something that's emphasized with Claude, OpenAI and Google or any AI company. We're always trying to figure out one, how do you get more context? How do you deploy agents on a wide, wide breadth of knowledge? How do you make the most of your subject matter experts? And how do you operate with enough leverage

and move quick? And then not only all of those things, but also how do you maintain thousands of agents? How do you think about compliance or consistent context protocols for those agents to be operating consistently across your organization? That's what Google talked about. And so their vision is really, "Hey, we understand that you're going to have thousands of agents one day.

We could be the platform to manage that. Doesn't matter if you're an AWS or you are a Microsoft shop. We could be your one place where you create and manage your agents. We'll give you context and APIs and integrations with MCPs, with model context protocols, where you can integrate wherever you want. You like Salesforce, you like HubSpot, you like

this app, you like Linear, Notion, and so on." You're able to connect all those apps with your agents and those agents are able to complete tasks on your behalf. And It's pretty cool if you are an enterprise. So just farewell to this episode, I guess, because in this episode I'd ran into so many issues.

Dalton Anderson (02:30.698) So first thing is I had to sign up for Gemini enterprise to look at what was demoed or some of the stuff that was demoed. The other stuff is in Google Correct which honestly not that comfortable sharing live in a podcast. Like I feel that I might accidentally share something I'm not supposed to. So I'd rather just not share anything at all.

And it's an active production environment, so I don't want to mess with it. The second thing is I had to sign up for the Gemini enterprise platform, which I personally was confused because I have the Gemini Ultra subscription, which I think they changed to something else. So there's just a whole bunch of weird stuff going on with the subscriptions and the platform. And then when I signed up for the 30 day trial,

creating a demo for this episode, I got rate limited. And so I don't have any actual agents to demo. And then on top of that, I guess my camera, my new camera, you could see me in all my glory now. I've got a nice DSLR Sony camera, one of those typical podcast or talking head video cameras. and you could see me, I could see you or you can, you could feel like I can see you.

And I don't know what happened, but apparently the cord that was connected to the computer that has some instrument that translates the camera's video into something that the computer can understand, it got caught on something and basically ripped out. And so the USB link, it's called a Cam Link, was messed up. It just was. And I had to fix it because it was

like a triangle instead of a square. And so I had to get pliers and bend it and put it back together. And now I've lost the USB port on my desktop computer, which is not ideal at all, but at least it all works. And then there's just so much stuff going on with this episode. Regardless, I will be referencing my phone when talking through stuff because

Dalton Anderson (04:58.315) can't remember everything and I don't have a printer because I live in New York City and my place is super small and I'm not buying a printer just for you. So deal with it. Okay. So in this episode, we're going to be discussing Cloud Next and I'll have some clips of stuff that I think are worthy to share. They had two live demos and then they had some notable information in the beginning of the video. If I had a choice and told you what should you watch, I would watch like the first 15 minutes of

the video in the beginning. And then I would just watch two of the live demos, which I thought were quite interesting. Or you could just watch this episode or listen to it and you'll get the gist. Okay. So the first thing that I wanted to share was a quick snippet from Sundar. and it is the

Next opening share with audio. It is their investments, so their CapEx. Their capital expenditure.

Dalton Anderson (06:36.237) $185 billion in CapEx from, I think, from 35 billion is a huge, a huge difference. Like just the scale of what $185 billion is hard to understand. So but that was interesting how much they're spending in infrastructure, research, R & D, et cetera.

And that really goes into what is the process of what they're working on and how is it going to work? And so they explained it in layers. So there's multiple layers within this. And basically what they're emphasizing is, "Hey, we're the platform that provides the context, the data, the execution, the governance and the consistency." And to do that, they divided

they divided the, I would say like the instruction protocol and information into different layers. So there's a security layer, there's a governance layer, there is a context layer, there's a data layer. And within those layers might have different components, like context might have your email, your drive, your Google Drive, your OneDrive, your Outlook, those kinds of things. And then,

your data layer might have your database, your database views that are specialized for agents to access because you don't want to give your agents read and write read and write access to probably any database, but he especially don't want them reading from your production database because it's going to slow everything down. So saying all that to share my screen and show the stuff instead.

Dalton Anderson (08:44.941) Okay, so what they're saying is agentic task force. Task force. It is is what it sounds like. It's for agents to be able to complete tasks. And then there's the agentic platform and models, which is Gemini and or other models that they're integrated with like Anthropic, although they own quite a bit of Anthropic, currently 16%, but then they have additional investment. And I think

that the total equity would go up to 20 % or something like that. It's quite a bit. It's more than people think. And they also supply all the compute or majority of it for Anthropic. So basically there's a subsidiary. They can do their own thing, but they're quite similar. And then there's the agentic defense, which prevents your agents from executing things that they shouldn't. And

that's done twofold. One, from executing and collecting information that is external to your organization. And then the other part is if somebody got into your organization, they won't have access to create agents that could access things that they shouldn't. So each agent is typically, not typically, each agent is set up in a way to where everybody is treated like an individual user.

And so they all have agentic permissions and those permissions are generated and are consistent. And so if you were to create an agent as a random user, like if I hacked into your company and I try to create an agent to access personalized data or whatever, it would get flagged and it wouldn't be allowed.

And if it was allowed, it would also get flagged by the security agent. Like there's this weird IP address that isn't typical and we need to execute something right now and it would escalate it. And we can close down. There's a full demo on that on here, the security and the red team, the green team and the blue team, which is really cool. Honestly, it was a great acquisition from Google to acquire Wiz and the thing that they rolled out in the live demo was great.

Dalton Anderson (11:10.638) I don't know. I thought that was incredible, honestly, but I'm not a security expert. So, but I can get the gist of it. It's basically AI agents looking through your code base and your security permissions nonstop, suggesting what vulnerabilities are created being your internal hacker and then mimicking basically what teams would do consistently, but since

code is being generated so fast, people can't keep up with how fast the code is being generated. So you need to then create AI security engineers. And that's what Wiz is created slash orchestrated within their platform. Okay. Anyways, So the agentic defense, and then there's the agentic data cloud, which what I was talking about, where you can integrate data from AWS, Microsoft Azure, and then

of course, Google Cloud with BigQuery and buckets. You can do all those things. No problem. And then the AI hypercomputer. To my understanding, it's the physics based model to ground everything within reasoning? It's typical for these other models to have a lot of their foundational truths based in physics. Because

it makes sense. Like if you can understand physics, you should be able to understand other things because then you have enough critical thought and reasoning to understand other aspects of the world if you can understand the ground truths of physics.

So that was that section. Let me check my phone real quick. Might have to mark this as a edit, but.

Dalton Anderson (13:05.197) Okay, so the way that they're doing all of this, this is the back end, but then the front end of this is going to be a normal.

Dalton Anderson (13:20.397) Where is it at?

here we go. So that's the back end and how everything's going be done. But what will users see on the day to day? The user will see something like this and the user will see this interface where it makes it easy to create agents. And so there's a section for agents to be created. And then there's chats. What I really think that this and this has just come out, by the way. I haven't been able to utilize it that much because one,

I'm not a full enterprise, so I don't have as much data and things going on. And it's hard to simulate all that. And then, I think you have to create agents within here, within your backend for you to get full use out of creating agentic workflows, because one thing that they talk about in detail is agentic orchestration and agent to agent protocol or A2A.

And basically it's a way for agents to communicate with other agents and then hand off tasks, which isn't too crazy, like it's been around for some time. But to do it at scale and the way that they did it in their demo, I thought was great.

So this is what mine looks like if you created a new agent. I can't right now, unfortunately. Google, hello. You are giving me a rate limits right before the episode. That's disappointing. I only created five agents and the ones I did create were pretty bad. So I was trying again and it would have been nice to have higher rate limits. So I created two agents and then I didn't even save them. I think they're, they're still drafts.

Dalton Anderson (15:09.805) And then I was trying to create three more just to test things out. Although it didn't turn out very well. So I think that there's just a lack of context when it comes to creating the agents, because I just don't have that much information on my Google Drive. A lot of my stuff's in the code base. And this isn't what This topic of the episode is not about my code base because it's private.

Okay. So one thing I did think was really cool before I get in the demo of what they showed within Google Cloud Next is you can ask it about your workspace, which you can't do natively with Google Gemini because it doesn't know everything. And if you're on a company,

plan, it doesn't allow you to access that.

Okay, so it has all this stuff. It's like, "Okay, well, you've got some things from Uplift Desk. You've got some stuff from Typefully." Typefully is a great company. I went to the Intratech NYC or Intratech NY. Went to that event. That was a great event. Appreciate that, Intratech NY. A couple weeks ago. So shout out to them. But then it talks through, "Hey, I see in your chat that

you've got AI news, you've got this stuff going on, consistently publishing stuff within your Google chat", which is true. I do have that going on, which is nice. I like to read them all the time because it comes in every morning and I just plop through, read the weather, read any news about insurance or AI. And so it's quite nice. And then the next thing I suggested was a travel agent. It says I'm going to a wedding.

Dalton Anderson (17:03.021) "See you're planning a trip as indicated by the train tickets for a wedding in Spain." Look at you. Mr. Creeper AI, huh?

And then it also said that a podcast production agent, "I found an email thread discussing the rescheduling of a recording in a press kit for a podcast episode. We could create a podcast production agent." And I was like, tell me more about the podcast agent." And then it gave me all these examples. And I was going to create two of them and just test them out before the show. And then that's when I got hit with the rate limit. And I was like,

"I guess I got to pivot once again," because I was going to do something else before that. And it was just taking too long. I was going to create a agent for YouTube to read through my videos and provide marketing advice to see how I can make my videos perform better on YouTube, which I thought would have been interesting, but it was taking too long and

it wasn't special because it was already a pre-built agent within the agent platform.

it regardless. Okay, so let's pivot over to

Dalton Anderson (18:31.564) Man.

Dalton Anderson (18:46.478) All right, here we go. Share a tab instead.

Okay, so this is the demo that I talking about.

Dalton Anderson (21:24.728) I'm going to skip it ahead just a sec.

Dalton Anderson (21:51.535) So this is a demo of the usage of this MCP connection between different information, being able to bring it all together and accelerate decision making. And then there's another piece that I wanted to share

Dalton Anderson (22:15.899) here with an agent task force.

Dalton Anderson (22:44.302) Okay, hold on. Give me a moment. I want to get the demo. I had it pulled up earlier, but then I had doubts. Okay, So this is is stuff that I think is really important to emphasize. If you're a company and trying to innovate, you've got to think about yourself as the first customer. In the most innovative products, you are the first customer. And if you think about that in a first customer principle, when you're rolling out products, you're the first customer,

you're going to want to get the good stuff. You're not going to limit everything on, "We need to get this out. We need to get our Q1 numbers up for next year. The forecast is looking bleak." You'll take the time. You'll have the edge cases. You'll have the company in a good spot when the product rolls out. When I say the company in good spot, more or less the product, when it rolls out, it's already vetted.

And people are going to love it because you love it because you are the customer. You literally use it every day. So the problems that the customer would have, you have already solved. And that's something that Google emphasized during this event. They consistently referenced that they are their first customer and they provided demos of

partners and they also provided demos of stuff that they're doing in house, which I think is really important. And these are live demos, which is also important. I don't trust companies that don't do live demos. Sorry. You got to take the risk to get the reward. And so in this part, there's a live demo of what I think is something really good. And this is going to be

a service agent that they rolled out for YouTube TV.

Dalton Anderson (27:30.19) sharing this stuff. Such a baller execution of the demo. There's a couple things that you can see maybe here, but when the person's demoing, sorry, I think I forgot their name. I think their name's Patrick.

Dalton Anderson (27:54.092) It's okay.

Dalton Anderson (28:00.087) Yep, Patrick. There it is. Sorry. Sorry, Patrick. OK, So when Patrick's doing the demo, there was a couple of things that he did that were great. Within there, he said many different words wasn't necessarily super duper clear, but got around to the point. And then when the AI was explaining the offering, he had interrupted the AI and said something like, "Yeah, go ahead, whatever."

And then it sent the link and then he had said, okay, now, and it didn't finish what it was supposed to do. The idea would be to close the deal and get Patrick to sign up. But then he asked AI, the AI customer service agent to then, "Hey, can you translate this in Spanish?" And then it translates in Spanish well enough. I don't know Spanish, but it sounded great.

And then from there carried on with the conversation back in English. There's a lot going on there. And they demoed something back with Salesforce, with Sales Agent Force last year that was pretty cool. Understanding what somebody's taking a photo of and live screen, live sharing their phone to discover what plant they have and things like that. But this is

on another level because it's within Google's platform and it's rolled out. It's in production and they did it in a couple of weeks, I think, six weeks, I think they said. From idea to production, which is pretty fast, I feel like, to roll out an AI voice agent to millions of people.

Typically would take more than six weeks, but they're moving at the speed of, you know what, I'm not going to say it, but you know it.

Dalton Anderson (30:10.862) But that being said, they had a lot of great announcements and I think it puts them in a really good spot when you're talking about foundational platforms, because I think there's going to be less of a gap on models. Everybody's rolling out the same thing within a couple of weeks. You might have a couple of week advantage with an Anthropic or OpenAI or with Gemini, but then the gap fades pretty quickly. And what you really,

I think, in my opinion, what really sets the difference is context and consistency. If you can get the right information to the agent, then the agent can do more. And if you can consistently get the agent to execute, which both is compute and governance of the agent.

If you can do that consistently, plus with contacts like your emails, your calendar meetings, your database, all that stuff, your search history, you've got a pretty good moat. And if you're banking on the models,

being the same across the board and really what you're competing on is context availability and compute context and compute, compile the context and compute availability, then you're well positioned to win the race. I've always thought Google was going to the race. So, hey, have at it, but they have a lot to do. I think they're still far behind in certain areas like

coding compared to Anthropic or Codex. But they're getting there, especially with their Antigravity editor that they rolled out, which is, I think, really good. Way better than Zed or VS Code. I did not like it at first, but that was because it was more of just a simple fork of VS Code. And I was tired of forks of VS Code. Like, "Come on, let's build something from scratch and use less resources.

Dalton Anderson (32:30.092) Let's get something more efficient."

But when it comes to productivity, Antigravity is legit.

Dalton Anderson (32:41.3) All of that being said, I think Google is becoming very well positioned with this new platform and their approach. It puts them, like probably in six months on another level. I would think if they keep moving at this rate, six months from now, they should be in a very good position because they have Chrome. So they have all the search browser stuff,

web crawlers. They can do that at minimal costs because they already do that consistently for daily users. So what's the cost to do it for just AI? They have YouTube, which has a whole bunch of videos, social interactions, trends, also used to enrich search history, not search history, also used to enrich.

search and for AI to create videos. They have short form. They have long form, both of which TikTok and Meta really don't have key availability on. They have a massive amount of talent. They have currently the most AI talent in the world with both their Google employees and Google DeepMind

Dalton Anderson (34:09.617) So, oh, and They also have the majority of AI compute. And then they recently just rolled out new chips, the TPU chips. And this is the first I've seen that they have both chips for surfacing and for training. So they're going to have training chips and then they're going to have serving chips because they do different things. Which makes sense, but it shows how serious they are to develop two separate chips

to do two separate tasks very well instead of one ship that can do both tasks but just not as good as if you developed single chips. I wonder what the margins are between one single unified chip architecture or two chips, one for training, one for servicing of AI models. How much better is it? I don't know. They didn't say, but I'm sure it is, or they wouldn't have spent the money.

They did the math, I'm sure. Hopefully. You would think. You never know. Things like this happen. I've seen it before.

Dalton Anderson (35:16.982) But the Google Cloud Next is just the antithesis of everything Google was committed to. Google said a year ago that they were going to do everything they can to change the path that the company's on and that they will be serving and part of the AI revolution and they will not fail. And so they've been shipping at an incredible rate. That, and I think in the xAI or

just shipping very fast. Whereas OpenAI and Anthropic are really well positioned. OpenAI has a massive mode when it comes to third party integrations. Anthropic has a good corporate base and great UI and has Cowork and Claude Code so they have a great base as well.

But Google doesn't really have a base. They just have potential at the moment. Like I don't know anybody that is shouting from the rooftop like, "I love Google Gemini. Gemini is amazing." I like Gemini because I think it's cost efficient. and I like how big the context protocol, like not, context protocol, I didn't say that. I like how big the context.

window is how many tokens I don't know I'm blanking out I like how much or how big the context window is I don't know I like how large the I don't know This is what I like about Gemini. I like that their model can handle a lot of tokens

and just it allows you to have more complex conversations or inputs.

Dalton Anderson (37:21.43) That being said, I was going to show you something cool that I was working on in my free time. was trying to build a, let me do this without leaking my email or emails. I was going to demo a comparison between the two, but I got rate limited, unfortunately. So you don't get a demo between two of them. You just get a demo between one.

Dalton Anderson (37:54.606) Where is it?

Dalton Anderson (38:03.818) studio.

Dalton Anderson (38:10.7) Hmm.

Dalton Anderson (38:14.835) There it is, it just took forever to load. Okay, so I'll show you.

Let me share my screen.

Dalton Anderson (38:35.468) All right, so this is the studio. This is the workspace studio and this is the non-enterprise version, but basically it's Zapier on steroids when it comes to things within Google. Things outside of Google, it doesn't really do that well, but if you're within Google in their suite it does well. Similar to Power Automate from Microsoft.

I've used both those tools. I think Power Automate is probably more powerful, although Workspace Studio is easier to get started. if you're within the suite of Google, I think it's better. But to each their own.

But I set up a flow and this flow is for podcast guests. So one thing that comes through is I'll get an email from a guest and or potential guests and I get quite a few emails. So I need to know one, I need to know is this email related to the podcast? Okay, if that's true, let's move forward with that.

And then is this episode related to like sponsorship or is it related to a pitch or is it some alumni? Let's identify that. And so if it is true that it's a pitch within the decision to.

Dalton Anderson (40:18.39) If that is true, then it's going to ask Gemini to read through the email. wait, this is the wrong workflow. This is the automated one that I set up. Hold on. I was like, this doesn't look right. So everything I said was true, just not the steps I was showing. OK, so this is it. So we add labels. So with this workspace studio, you can have Gemini labels or AI powered labels, which are separate

from your email and then they create these AI labels or you help create them. And so they're less rigid than they would be if you created them. So it's quite nice. You can set instructions for each of these. And then we're gonna summarize the email and create labels. So the first thing I wanna know is: Is it a pitch? Is it alumni? Is it sponsor? Is it other?

And then once I know that information, I want to extract everything. I want to just take everything, rip it out of the email and put it in a structured format. And so that that would entail taking that email, looking at the email body, the people sent it, their header, the website, everything. Trying to glean as much information from the email that I can. And then I serve that to Gemini, ask Gemini to

read through all the context that I extracted and then understand the pitch and provide some context on whether or not this is a good fit given the historic guests that have had on the show. And then help me draft an email by putting in what they think the verdict should be, the brief and an email draft. They'll draft a reply, but one of the big things that I like

is that I have it set up to where it will write to my Google Drive all the information that we've extracted. And so it makes it really easy to understand what's going on. And then if you just wanted to test it. We could test it right now. We could do this when there's a random email from St. Patrick's Day Eve. So this is a email that is not a pitch.

Dalton Anderson (42:42.826) And I don't know what this is for, so, flight's nothing odd.

Dalton Anderson (42:52.578) Maybe some weird spam.

Dalton Anderson (42:56.908) Okay, so it is a person.

My bad, Patrick. A lot of Patrick's today.

Dalton Anderson (43:10.478) I though it was an email newsletter because it has an emoji in the subject line.

Author, founder, mentor.

So it runs through all these steps, the steps that I described earlier, and this is my agentic workflow. This isn't really agentic, but it's a replacement of robotic automation. Robotic, forgot the abbreviation for it, but yes, that. And

it adds it to this row within Venture Step. And then from there I could read the verdict. I can run with it. So in celebration of this demo and the episode, I'm going to turn on the automation flow. I did it. Keys to the kingdom, baby.

I really wish I could do a comparison demo between the two, but unfortunately I got re-limited and I've said that several times this show, but I am disappointed with Google for doing that to me. But circling back over to everything I talked about, Google's really showing up and all these companies are showing up. So I'm really happy that you have the opportunity, one, to be on the

Dalton Anderson (44:37.358) just the forefront of these things. It's just such an interesting time to be part of this. And then the second thing is stick to your platform. Like if you like OpenAI, use OpenAI. If you like Anthropic, use that, vice versa. Don't switch, just use what you're comfortable with and what you've got set up. You've made the commitment. Switching platforms is expensive and there are tools that you can switch.

with like, if you were changing your music platform, but don't switch because you feel like you need to because you're missing out because something that another platform has you'll have eventually. And so if it was a deal breaker, you probably wouldn't have been on that platform in the first place. So I would let it be, let it ride. So that being said, I've got the new camera set up. I hope that you can see me.

very clear and also you can see how tired I am. So funny. Okay. Of course, wherever you are in this world, good evening, good afternoon, good morning. Thanks for listening and listen in again next week. Goodbye.