Deep Search agent deployment and use

Lightly edited for readability: mis-heard words fixed and filler words trimmed. Words in [brackets] were unclear in the recording.

Welcome to episode four of Grant Builds With AI. In this episode, the end result that I now have up and running is: I deployed a Google Agent Garden agent called Deep Search. I deployed it, and it is currently running on the Google Cloud, somewhere. And I interact with it through my Claude account, through Claude Code. I send the message through Claude, who gives it to the agent, and then I receive a result back that Claude turns into a big Word document and puts in a folder on my computer. And it also gives me an executive summary of it.

All right, so what does that mean? Well, for my next project, I had been thinking about a couple of things, but one of them was that I wanted to start using the different tools available out there. I've always been a Google fanboy, so I made a new Google account, the Grant Builds With AI Google account. And my initial run at this was, I was going to copy what I had done with the D&D bot, but do it using Google tools. And that quickly fizzled out. It was pretty hard, and I was trying to use Gemini to do the cut and paste, and I just kept not getting anywhere with it.

But I did get a virtual machine up and running, a free-tier bot that I called Discord bot, but I never loaded anything on it. To do that, I had to interact quite a bit with the Google Cloud platform. I made an account and just started moving around within it. That was really interesting. In the beginning, I found it really hard, because nothing made sense to me. It reminded me of AWS. It looked like it was trying to be more user-friendly, but as I moved around in it, I could tell this looks like you should have a little more experience as you're setting things up. But I would take screenshots and send them to Gemini, and it would come back, and I kind of clicked my way around. I got my virtual machine instance up and running, and then stopped, and was like, huh, okay, I don't want to move forward with this. I think I was kind of tired of building a Discord bot.

So then you just start poking around on the Google Cloud website, which is pretty big. And you notice there's something that says "Agent Platform." I'm like, well, that's what I want to do. I want to work with agents, which I don't really understand what that means, but learn by doing. So you click in, and they have something called the Agent Garden, and near the beginning is an agent called DeepSeek. Not DeepSeek, I'm sorry, that's the Chinese model. Deep Search.

You can click on it. It says it's an agent that you pose a complicated question to. It digests it and sends you back a proposal of how it's going to do the research, and then it deploys about ten agents that start using Google Search and start searching through your question. They get a whole bunch of information back, and then it digests all that information and gives you a pretty detailed report. And I'm like, this sounds like a great idea. And right next to the agent, there's a button that says "Deploy," and I'm like, this is great.

So you just deploy it. And then what happens next? Well, it opens up the Cloud Shell. It opens up a terminal. And you're like, who am I talking to here? Okay. And it kind of pastes in some commands for deploying the agent: it installs the agent's connections, installs its code, points it in the right directions, and you're up and running after you hit deploy.

Well, no. It just immediately gives you an error, in the command line, in the Google Cloud Shell terminal. And I'm just staring at it, like, okay, I don't know what's going on here. I don't know what's happening. I don't even really know who I'm talking to when I'm working in the Google Cloud Shell. Well, spoiler alert: I still don't understand who I'm talking to in the Google Cloud Shell. In my head, I'm now talking to the Google Cloud. As I've read a little more, it seems that it's launching its own instance. And the end result is, you're trying to deploy the agent so it's living in Agent Runtime land, where it just gets assigned some resources and it keeps track of it, and it just lives there as part of your Google Cloud project. And it is tied to your billing account, and more on that, because that becomes important later.

Okay. So now I've just got these command line errors, and I'm kind of frustrated working with Gemini. I'm like, all right, well, let's see what Claude can do on this. And again, this is another time where Claude just kind of took over this project for me. I would copy-paste in the errors that I was seeing in my terminal session on Google Cloud, and it's telling me, "Oh yeah, no, we got this. We just need to rewrite the code of this agent, because, yeah, Google has Agent Garden, but none of those deploy buttons work. They've been changing things, and they all point in the wrong direction, and they're using the wrong model. But don't worry, we can tweak this. We can make it happen." And I'm like, okay. So it starts feeding me some command line things that I'm putting in the Google Cloud terminal.

And even in the beginning, it goes, "If you like, I can take this over for you. You could set up permissions where I can just start working for you in Google Cloud." And I was scared of that. I'm like, no. Because I also know that with the Google Cloud billing account, there is no way to set a max bill on it. You can get alerts as you start to accumulate cost, and you can set hard limits on how many resources it can call: how many CPUs, how much memory, whether you can link it to GPUs or TPUs. And I'm like, no, no, Claude. So we just keep cutting and pasting, moving through. It keeps not working. It keeps not working.

And it's starting to make sense to me why. This is another project that Google put out there. It doesn't work. It's not push-button turnkey. But if you've got experience as a developer or in software, then from their side, the error messages that they're giving you should make perfect sense to you. It should be, say, "Hey, don't worry, we haven't updated this, but you can go manually update it." Or, "Oh yeah, that points in the wrong direction. We renamed it." Something like, instead of something, there's a Vertex something now. "And oh yeah, don't worry, our CLI does this now." Well, Claude understood all that. Not the first time, but it would just keep failing, giving new error messages, and then all of a sudden it works, and Claude tells me, "Yeah, your agent's deployed."

No! What? It's like, yeah, you can find it here, because it's not really clear where it goes. So you go back in, and just like your virtual machine, you have to click through to find it. Hold on, we've got a car coming. Yeah, so you click through, and you can see that you have an agent deployed. And where is it deployed? It's deployed in Agent Runtime land, which, again, is just hosted on the cloud somewhere. But it's up. It's running. It's billing my account. And it's just sitting there, waiting to be asked a question so that it can start deploying its agents.

So how do you talk to it? You have to click through. How do you talk to it? And Claude goes, "Oh, well, now you have to build your interface to talk with it." I'm like, what? Not exactly plug-and-play. So there are a couple of options. One of them is running a website, another program that you've designed just to talk to your agent, which kind of makes sense. I'm like, "Claude, give me some more options." Because there's a playground setting where you can click through and ask the agent a question, but you're interacting in a text-only window that's maybe a fourth or less of your screen on the Google Cloud platform interface. And this thing is sending back a 62-page research report of how it did it, its results, its summaries, and it's really not practical to interact with it and read that in that little text window.

So I'm like, okay, I guess I've got to build a web interface for this now. And then Claude goes, "Or, hey, remember, you can just log me in, and I'll take over your project on Google Cloud." And I'm just like, yeah, sure, let's do that. Because this is where I do give myself some credit: I set up a whole separate Google account for this, so I'm not using my personal Google account. I made a whole account for Grant Builds With AI, grantbuildswithai@gmail.com. You can email me there.

All right. So that has to happen in the terminal sessions on my computer. So Claude gets all logged in on the project, and then here's the end result, and, quote, this is where the magic happens. I think I'll have to go back and review exactly what's happening here, but let me tell you what the end result looks like for me. It's in a Claude Code session, which I've got labeled and clearly assigned to its own local folder, and it's got memory that it says is going to be persistent for how it's logged in. Although it says the login is going to expire eventually, and I'll need to do that again.

I ask the question that I want. The first question I asked was: "Please compare tirzepatide and retatrutide in relation to lean muscle mass, muscle functionality and fitness, and also in comparison to bariatric surgery. Please focus only on peer-reviewed articles." I put in a little bit about where to search, but not really. I just kind of wanted to see where it goes. I should have put this in front of me, but I sculpted a message to ask the Deep Search agent about something I was sort of familiar with, so that I could gauge its output.

So I put together a question, but I do this in a Claude thread, and then I go, all right, send it. It sends it. The agent sends back, and this is normal, an outline of how it's going to do the work. It goes, "Okay, here's how I propose we do that." We actually made a couple changes to that. And then you hit run, and you wait about five, six minutes. The agent does its run, and now the agent has returned a 62-page document of its report. A lot of it is just the logs of how it did its run, but there's also a pretty strong 20-page findings report, with a summary.

Now, all that's being presented to Claude, inside Google Cloud. Claude then takes it and makes a couple of separate documents: one is the run log showing how the agent did it, and then it makes a Word document of the findings and summary. And then it just gives me a little executive summary inside the chat and stores all those in a folder on my local computer. And it works. I'm really impressed with the output on this. I'll post it on the website. I'm like, wow, that worked. Claude became my web interface for this, to talk to the agent.

Then I sent another question, asking it to do a deep search on rumors or confirmed news about Google's newest model releases. Because there's a lot of speculation around Google, that it has, quote-unquote, fallen behind in the model wars and is focusing on Flash, focusing on low-cost compute, releasing a lot of Flash models, when is Gemini Pro 4.0 coming, things like that. And again, it works. I've got another big answer on that, which I'll post to the website. It's kind of a snapshot, because it's as of this date. And I'm pretty impressed.

So now I'm at my end state. I have a deployed Deep Search agent that lives in my Google Cloud project. I interact with it through Claude, and it outputs a pretty nice Word document of the work that it did. I also downloaded all the code of the agent onto my computer, and my next project is to start tweaking this agent and rewriting its code a bit, to see if I can get it to point to different models. Although it's already running into the fact that only the Google models can use Google Search. But you could use a different model. It looks like Google actually has an agreement with Claude. Can I turn on the critique part of the agent? After the Google models have gone out and searched Google and brought a whole bunch of information back, can the report be written by Claude Opus 5.5? Because I don't know, maybe it works better than the Flash ones. So that's kind of the next project.

This episode went long, but that brought up another thing. I guess I did build one more thing. As I started doing this, I could start seeing: you deploy an agent, what if it just starts doing crazy stuff and billing my account? So I didn't ask Claude, but I started looking around in the Google Cloud platform, because all the other things that I've worked with have usage caps or spending caps. You can basically push a button that says, after I've spent 30 dollars, shut my account down. Well, Google does not have that. It has a notification, and like I said, it's got usage caps. But I'm sitting there thinking, if my account got hacked, somebody could basically start mining Bitcoin, approve more usage, and keep mining Bitcoin.

And I did a little research, and I think I'd heard about this before. This happens. Google's got a kill switch thing, but its billing reports run four hours or so behind its usage. And so there have been cases of businesses getting a bill from Google Cloud for a hundred thousand dollars because their cloud account got hacked and started mining cryptocurrency. So I'm like, I don't want that.

So I told Claude the problem. I said, I don't like this. And Claude goes, "Oh yeah, but don't worry. Why don't we just build a program that sits on the cloud and monitors a message channel within the Google platform?" And we set in our settings that the billing office sends a message through the channel when you've reached, let's say, 40 dollars of spending, and then this little software sitting in the Google Cloud reads that message. It's basically, whenever there's a message, the little program looks at it and goes, is it above 50 or below 50? And if it's above 50, it just shuts it down. It says, nope, can't spend any more. Stop using everything.

And I'm like, yeah, that's exactly what I wanted. Why doesn't Google offer this as a feature? Actually, Claude had some good explanations: this is more focused on enterprises, where downtime can be a little more catastrophic, and serious enterprises should be more sophisticated and have their own monitoring. And some of it made sense. So yeah, I built that too. I built and deployed a feature that puts a spending cap on my Google Cloud. So fingers crossed on that one. That would be the end of Grant Builds With AI, if my cloud project got hacked and I got a six-figure bill from Google.

Hey, FYI, universe: yes, I turned on two-factor for my Google account. And I think my next plan is to ask Claude for a check on my security setup every once in a while. And I did that. I said, "Hey Claude, this is why I'm worried. Can you run a check on my security?" Actually, I did that yesterday. While messing around, I had generated an API key, and Claude goes to use this API key, and I'm like, no, delete it. Delete it. And it said, "Good job. That was one of your vulnerabilities. If somebody had gotten that, they could do stuff with your cloud platform." Which totally makes sense. And it dawned on me: how come I had that just sitting out there?

All right. Well, this episode has run 22 minutes. That's a wrap. It's a nice walk this morning. It's gorgeous. We'll see when we get this published. All right, have a great day. Bye.

Oh, I should give a shout-out to Will Tuck, who I had a discussion with about some of these things. I really enjoyed it. Will's got a lot more experience in this than I do, and he gave me some encouragement, so I really appreciate it. All right, have a great day, bye!