Build ANYTHING with Gemini 3 | The Agent Factory Podcast

[music] Hi everyone and welcome to the agent factory the podcast that goes beyond the hype and dives into building production ready AI agents. I'm Smitha Cullen

>> and I'm Vlad Kalisnikov >> and today we're joined by Brandon Hancock who is a full stack engineer and who is also known on YouTube as AI with Brandon where he teaches AI to over

80,000 developers. Welcome to the show

80,000 developers. Welcome to the show Brandon. Hey Smith. Hey Vlad. So excited

Brandon. Hey Smith. Hey Vlad. So excited

to be here today. Awesome world of everything that's going on right now with AI and I'm so excited to dive into it with you guys today.

>> Yes. Uh it is a very special day. We are

recording just a few hours uh after Google released two new products. birth

our new flagship model Gemini 3 where we had a lot of anticipation for that and a surprise anti-gravity um our agentic

coding environment with lots of features uh to explore and um in just in general it's very exciting and um we will show

you a few demos um of Gemini 3 obviously uh Brandon has a very special demo himself uh with Gemini CLI And I've got some surprises, too.

>> Awesome. So, let's kick it off with Gemini 3 Pro. Uh, I've gotten the chance to play with it already, and I've been super impressed. It's really great at

super impressed. It's really great at advanced highle reasoning, and also, it's able to uh complete complex tasks really well. Also, it is able to do

really well. Also, it is able to do stuff like agentic operations really well. So, it makes it really the ideal

well. So, it makes it really the ideal model to build AI agents with. and I

spent a ton of time already trying to build AI agents with Gemini 3 Pro and I've been blown away. Uh, one thing I've really liked using it in is actually AI Studio. So, let me actually hop into AI

Studio. So, let me actually hop into AI Studio and showcase what I've been doing. Okay, so I'm in Google AI Studio

doing. Okay, so I'm in Google AI Studio and what I've been thinking recently is about building a personal website. I

know I'm I'm a dev at Google and I should probably have one, but I don't.

Uh so what I've done is actually I took my LinkedIn page and I exported it as a PDF. I also used Nano Banana to create

PDF. I also used Nano Banana to create an inspiration picture of what I want my website to look like. And then I'm

uploading all of that into build in uh Google AI Studio to help build out that website. So let me start off by

website. So let me start off by uploading the PDF of my profile from LinkedIn.

And then I'm also going to be uploading a profile image of me so it gets that right. And um I'm also going to be

right. And um I'm also going to be uploading that inspiration picture of how that website should be looking like.

Okay. Uh help me create a personal portfolio website based on the profile.

PDF file which contains my LinkedIn profile and then use the inspo.png as inspiration for how the website should look like. And then make sure that you

look like. And then make sure that you include that profile picture of me as well.

I'm pumped to see how this turns out. I

would love to see some of the like initial pictures you did with uh Nano Banana because seriously one of the coolest AI image generators I've seen. I

would love to see what you did for inspiration cuz um would love to see how AI takes that input and does something with it. First, I used uh Gemini to give

with it. First, I used uh Gemini to give me a prompt to give to Nano Banana based on, you know, what I would like. And

using that prompt, Nanobanana came up with this really nice inspo pick which is pretty close to the website I'm looking for.

>> That's awesome.

>> Look at the speed that it's working at.

Um I one thing I noticed with Gemini 3, it's how fast it is.

Um, yeah, that that really makes a difference for me as someone who likes to iterate over and over on the same thing. Um, and yeah, I've noticed that

thing. Um, and yeah, I've noticed that it works really fast despite the fact that so many people are now on it and testing and trying and uh, you know,

offloading like huge data pieces on it and just seeing how it performs. Our TPUs are must be on fire, right?

[laughter] And speaking of being on fire, I really love all of these fire suggestions that AI Studio gives. So, kind of like, oh, if you want to add video understanding,

I can easily try to click this and then add it here so that it becomes the next prompt that I want to add to my website if I if I wanted to.

>> Less than 3 minutes in. Oh my god.

>> Looks pretty cool.

>> The image is no longer available.

>> Okay.

>> See, it's AI. You always need to uh look what it's done.

>> I mean, it it got all the details, right?

>> I love it. I love it. Look, it really got like I've been to your LinkedIn page, so I know what's there.

>> Uh it really put your work experience and email and yeah, you were a great public speaker, de machine learning, everything.

>> Yeah. Uh okay, let me actually try to fix that image. Right. What I really like is this annotation feature in uh Google AI Studio. So, I'm going to click this

>> and it allows you to annotate stuff that like is missing. So, for example, just this section and I can say um add that

section to chat and fix my image with the actual profile picture. So, while that's going real

picture. So, while that's going real fast, Smith though, one of the coolest things that I've really liked about AI Studio recently is its ability to like single shot. I've tried to like test it

single shot. I've tried to like test it as much as I can on like everything from like web applications all the way to like like senior year of college like level math and computer science and

electrical engineering. So, I like

electrical engineering. So, I like tested it like on the full gambit and it has crushed it on all levels like all the way from regular software to like full-blown like calculus 3 level math.

Like I know they say like smartest model ever. So I was like I'm going to try and

ever. So I was like I'm going to try and test it out and it is. It really is. Um

yeah absolutely been amazed. I love this image and and see how it also like originally it wanted to style it with the kind of black and white. Um yeah

because that's what your initial style was and it it kept that that thing. I I

love it.

>> Yeah. And then also the fact that I could have I gave that prompt very specifically to just this section and it didn't change the entire website because that can happen too.

And that's my demo and this is my website now.

>> Have you got to try publishing it yet out of curiosity in the top right?

>> Oh yeah. So the cool thing is that I can deploy this super easily to Cloud Run.

It's it's super easy.

>> Yeah. It just it just goes straight there. Uh takes a few minutes and you

there. Uh takes a few minutes and you have your uh website up and running. You

literally don't need any knowledge about Google Cloud or Cloud Run. Uh you click one button. Sorry, that's just one of

one button. Sorry, that's just one of the craziest parts that every time I keep working with more and more of like these uh tools from Google, it's like the barrier to entry to just build and ship stuff is insane. So, like in a lot

of my communities, there's people that are like, "I've never coded an app before, but I have an idea." And then they'll just hop into, you know, any of these tools and actually create something. Like, it's crazy how many

something. Like, it's crazy how many more people are now able to build real world apps uh thanks to these tools.

Like, it's it's always so exciting to see people light it up for the first time making their own apps.

>> And hey, Smith has got a website now.

You can hit contact me and uh [laughter] and she'll fix it.

>> Okay, Brandon, you have something really cool to showcase on Gemini CLI for us.

All right. So, here's what I wanted to show you guys that I've been working on using a combination of Gemini, Gemini, uh, CLI, and I'm actually going to show it a little bit in anti-gravity because

I'm pumped to show everything allin-one.

So, quick background, guys. So, what I'm trying to do now in the age of AI is I want to be AI first for everything I do.

I swear there's never enough hours in the day. So, what I'm always looking to

the day. So, what I'm always looking to do is to build out what I'm calling like AI employees to help out with task in every part of, you know, my YouTube business, other businesses. I want AI to

do as much as it can so I have more time to focus on the big things. So, what I really want to show in this video with you guys today is how I go about

building a market research AI employees.

And uh just super quick background just so that we're on the same page. We have

a um me and a co-founder have a startup where we're helping out emergency medical services like ambulance companies, fire departments. We're

basically helping them out with their SOAP reports. So we have a phenomenal

SOAP reports. So we have a phenomenal idea of exactly who our customer avatar is, but they're all across America. So

it's very hard to like, you know, I could manually spend the rest of the week on Google finding everyone in my state. But we have amazing AI tools that

state. But we have amazing AI tools that are using Google search. So I could easily at scale kind of search the whole market. So um this is what I wanted to

market. So um this is what I wanted to break down for you guys today. Basically

I'm pumped to showcase how we can actually use Gemini CLI which is one of the best CLI tools with you know having the ability to do Google search and thanks to you know the wide array uh

wide variety of models. We have

everything from you know Gemini 3 Pro when we really want to think about it all the way to Gemini 2.5 flash when we want to kind of do things at scale. and

Gemini CLI they have like a super generous free tier. I think it's like a thousand requests per day completely for free. So that's a like hey gotta love

free. So that's a like hey gotta love free. Um so yeah so uh long story short

free. Um so yeah so uh long story short absolutely using Gemini CLI to build out all of my AI employees. And one of the things that I want to walk you guys through is how you actually can do this.

Like how how on earth do you build an AI employee? And it's basically just done

employee? And it's basically just done through markdown files. So, um, so that's what I want to walk you through.

And just a quick challenge for the audience after you guys kind of get to see all this in action. I would love to challenge you guys, too, to try to build out each week like a new AI employee because as time goes on by the end of the year, like you're going to have a

full-blown workforce of Gemini models running around doing all the work for you guys.

>> That's awesome.

>> All right, perfect. And real quick, high level, how do we think about building AI employees with Gemini CLI? The way I like to think about it is we have the ability to use multiple models. So

here's what we want to do. Just like as Vlad pointed out, we want to kind of build out standard operating procedures, which are instructions, which is awesome because we get to use tools like Gemini

3 Pro, which are these, you know, insanely smart models to help write out well doumented instructions such as like, hey, please go search the internet. Here's the type of questions

internet. Here's the type of questions we want to look up. Here's the ultimate goal that we're trying to uh achieve.

And these models are phenomenal at writing very um specific and clear instructions. Then when it comes time to

instructions. Then when it comes time to do not, you know, not the highle thinking, but to actually go into like what I call like workerb mode, hey, fantastic. We can fall back on some of

fantastic. We can fall back on some of the more fast, efficient, and more affordable models like Gemini 2.5 flash to actually go off and follow these instructions. And the coolest part about

instructions. And the coolest part about Gemini CLI is you can actually spin up like you could really spin up a hundred different agents in parallel to search across all America. Like this is you

really have agents at scale. So the core principle of building out AI employees comes down to what these documents include. These are standard operating

include. These are standard operating procedures where basically we once go we once go through the process and clearly define what our inputs are, what task we

want to accomplish and then at the very end we describe the actual output that we're trying to do. We can use Gemini to go forth and follow these instructions.

So what I can say is um I can say please follow the research uh let's actually do this the research city instructions for

and then you can actually like in my case I have multiple cities that I want to go look up. So I can go through this process and just like grab a few cities and I can say for these

uh cities and then what's really nice is I can actually pass in the cities and then finally after that what we want to do to like maximize our agent capabilities is to actually run some

Python scripts to actually kick off Gemini in parallel because that's one of the coolest parts that I don't know if a lot of people know about Gemini is you can actually kick off Gemini in the normal editor or you can do something

like this where you can pass in a prompt which is a cheat code because you can basically kick off hundreds of agents in the background using this strategy. So

in this case we have some nice little um some nice little Python scripts that will go through the process of kicking off Gemini in parallel to actually go off and research everything that I need.

So we're just going to say use uh we'll sorry about this. We'll say use at process cities. And you know, at this

process cities. And you know, at this point, what it's going to do is we're now kind of kicking things off to where it's going to go through the process of kicking off multiple parallel agents to

go off and search in each one of these different cities in Miami to find potential customers that we could reach out and help with our company. Now, at

this point, what's nice is we're actually having AI call a script. So the

AI is going to understand the parameters of the script, pass in all the cities and kick off the workflow. So I know this kind of probably got a little deep in the weeds. Is there any questions? I

want to take it high level too just because I want people to understand um you know that you could pretty much build have agents do anything. That's

the cool part.

>> The scripts that you're talking about, is that something you hardcoded or Gemini CLI did that for you?

>> Great question. I don't code anymore, which is the funniest statement as a software engineer. I never code. I

software engineer. I never code. I

review I review a lot of code but um all I have to do is say hey Gemini CLI what I'm trying to do is I want the ability to give you and in this case I want to

give you three or four cities to go off and research and what I want you to do is run you know Gemini-P as a prompt and basically just paste in this type of uh

city right here or these type of instructions which is hey go research a city do it for this city uh and basically just kick this off 10 times or 10 20 times, however many cities I give

you. So, no, that's all done based on uh

you. So, no, that's all done based on uh Gemini of what I need that question, too.

>> No, not a question, more like a comment.

Uh last time I've been at the podcast, this very podcast, I also used this uh what I think is a cheat code of creating

uh you know little Python tool and passing it to Gemini. And first like as a as a software engineer even though you don't write it uh you you have this

ability to kind of prompt it properly and guide it to create uh those scripts and verify whether it's right or not. Uh

but the thing I love the most is that you don't have to explain how to use it.

>> You just give code and Gemini says like okay I look at it this I think how how to use it and 99.9% that's right about it. I love and even even if sometimes it

it. I love and even even if sometimes it uh runs it in the wrong way and the tool crashes, it looks at the error and then that next time it's right.

>> No, totally cheat code. Use it all the time. Yeah.

time. Yeah.

>> Yeah. I agree. The one other just quick side topic while this is a spinning off when it comes to the concept of building out AI employees, one thing I do want to mention is we basically get to use AI to

do everything if it has context of the problem. So, for example, um I also just

problem. So, for example, um I also just other other things, I use Gemini CLI as a ghostriter to help me respond to emails. So, you can easily hook up an

emails. So, you can easily hook up an MCP server to Gmail and fantastic. I can

just pass in information such as like, hey, here's Brandon's ghostriter. He it

knows how to write like me. Can you

please go pull down my emails and respond to the latest three emails?

Because it knows who I am. It knows how I like to write and it can just kick it off. Outside of that, I have to write

off. Outside of that, I have to write monthly status reports. Well, my agents has to write monthly stat supports. So,

um, key thing is just the more folders you can add into your AI employee directory, you just record the task and then have AI do it the next time. And I

I do this for if I'm ever doing something more than once, uh, I can guarantee you AI is doing it for me. So,

um, >> wait, so so was it your ghostriter responding to all of my emails?

[laughter] >> I No, no, that's funny. There is human in the

that's funny. There is human in the loop. I do I do review it before

loop. I do I do review it before everything goes out. But um I mean this is the coolest part guys like um we can now use AI to move faster. Um eventually

it's going to be my agent talking to your agent and it's just calendar events are going to appear up on on so hey we meet tomorrow kind of thing.

>> This is unreal. Yeah.

>> So while this is going I just want to show real fast kind of what's uh happening just so we can uh can move on.

But at the end of the day what we're doing is just customer research. So for

each one of these cities you can see it's starting to like save outputs. I

won't click on it because it's actually going to show like customer data, but it does a super deep dive into understanding exactly what's happening in that local city and then kind of just

taking the same principles. We uh what we get to do next is once again use standard operating procedures to take in all the information from these different cities. We aggregate it uh sorry I don't

cities. We aggregate it uh sorry I don't know why it's showing but we aggregate it into a database and then after that we have an outreach SOP that shows exactly how to write personal outreach to each one of our customers to see if

they would be interested in working with us. So this is kind of the workflow to

us. So this is kind of the workflow to where you basically just have the AI document the process and then repeat it at scale going forward. So yeah um seriously thank you guys for uh for

making uh making Jim and I work and all the new models. Absolutely having a field day with them. Automate everything

with Gemini. Huh?

>> That's the goal. [laughter]

>> Okay. So, just to recap, we saw Gemini in AI Studio. We saw it in Gemini CLI, which is a great agent, by the way. But

do you have another agent you want to showcase with us?

>> Yes. Uh I will show you an ADK agent.

So, an agent built with uh Google's uh agent development kit. And this agent is made with uh vio 3.1 obviously with

Gemini and like I said uh with ADK. So

what's special about this agent is it's specifically made to uh create educational videos and it it doesn't

just uh take like a script and uh um creates a video with a talking head that tells you what's written in in in the

docu documentation script. It takes uh a documentation page that uh you know if you think about uh documentation for developers it's full of code it's full

of uh bullet points uh you know numbered list etc. And it creates um somewhat engaging uh video where in this case uh

my favorite one of my favorite animals uh Capibara uh is telling you the story.

So um the agent itself actually consists of multiple sub aents right uh like we love it about ADK it gets this uh this

boss orchestrator of the wood agent that uh works with two sub agents one is the script sequencer the script sequencer does two important jobs first it adapts

the script so um when you know when we start with the documentation page we want it we want the text and that that story to sound natural and engaging. So

does the adaptation or uh the scabby bar of our character to sound natural and to not read you bullet points like you know like many speakers do of uh

presentations. Um second thing it does

presentations. Um second thing it does it makes sure that uh those sentences they're split in 8 second pieces and it's very important because you know you

know uh video models they don't produce very long videos uh they produce um short uh pieces and we want again those to sound natural and have nice transitions and that's uh that's the

second job that script sequencer had.

Once that is done um it passes those um sequences those chunks uh back to the orchestrator and the orchestrator calls video agent to generate videos. One

important thing and you probably saw it when I was scrolling is that uh for for the video generation I use nanobanana to

uh generate multiple views of my character. Right? I've got four views

character. Right? I've got four views and I ask the orchestrator to uh randomly choose a view for every uh

video that it creates. And when I pass a script, it along with the chunking uh it assigns um a view number to every chunk.

So the video is not just like a monotonic talking head in the same kind of think about this number one uh position uh but it it goes randomly between views that the video stays

engaging. Even if you use it like as a

engaging. Even if you use it like as a small window uh in a larger presentation uh you want some changes for that and

that actually makes a big difference.

Along with that um I uh provide um examples of prompts how they should um should be and should sound. And if you

look at the this prompt, you will see how I provide very detailed character uh description and voice description and the visual appearance of the character.

This is all very important for uh those videos to be consistent even though we have those uh initial you know pictures the start frame for the videos. Why?

Because um take this um you know view number two. If the view number two zooms

number two. If the view number two zooms out a bit, uh we want the clothing and the you know the other other parts of the body of copy bar to look the same as

as you know as a view that starts uh as a video that starts with number four. So

those pieces are important and uh as you know the video models they sort of like work without context. They didn't know what the other videos it just made. uh

uh all those instructions are extremely important for um character consistency and environment consistency for that

matter. So basically what I do is I pass

matter. So basically what I do is I pass uh this whole prompt with the script. I

pass the the four views that I generated using nano banana. Uh and then it goes through uh

banana. Uh and then it goes through uh the sequencing and through calling uh video agent and uh sorry yeah the video generation agent and you can see that

every time it generates a piece it says like oh video number one this is the link and this is the piece of the script that um it's it's going through and we

can scroll down and see that it generated 92 uh videos for me but it again automate everything Uh I got the list of all 92

URLs. It still requires human in the

URLs. It still requires human in the loop. Let's be honest like the videos

loop. Let's be honest like the videos that sometimes not perfect. So I watched I personally watched every single 8-second video. Um and uh then I didn't

8-second video. Um and uh then I didn't like some videos. I didn't like uh video number uh 49 because there was some like weird glitch that view um made and I

have to regenerate it. it regenerated it and I didn't like it again because it sounded weird and it asked to regenerate it again maybe just like a use different view number and it did it again and I

got those um all those views and then uh what I did next uh is um I used Gemini CLI to join those videos together right

again all I did I just like copied this list and I copied the two file I asked Gemini CLI use fmc to join videos

together. And here u it did it. Let me

together. And here u it did it. Let me

switch to uh my IDE.

>> That has to be some sort of like video editing hack to use Gemini CLI to splice together all of your videos.

>> Yep. So this is my anti-gravity. Um and

uh I'm actually working on another version of this agent. Uh you can notice that I I have script adaptation as script litter at different as different agents. Stay in tuned for for that

agents. Stay in tuned for for that release. Uh but I I just wanted to show

release. Uh but I I just wanted to show you like the kind of the structure of the agent.

I have this uh main agent with a huge prompt how like it should read the results of different uh sub agents and how it should actually instruct them and

I was adding that uh human in the loop moment when I wanted to wait after the script adaptation and uh let me sort of

like make edit edits in the script. Um

what else? Yeah, and to make those changes actually that's why I wanted to show it. I use anti-gravity. So I just

show it. I use anti-gravity. So I just opened my the previous version of of um this engine that I built using uh just

Gemini CLI and I asked anti-gravity this is what I want want to add and yeah it uh it did all the work for me. It

adjusted the prompts. It removed all prompts. It uh made changes in the agent

prompts. It uh made changes in the agent itself and sub aents. I'm I'm super pumped and I just see how it uh improves

my uh productivity primarily because and I that's the first time I'm I'm seeing it. I prefer working in IDs over CLIs.

it. I prefer working in IDs over CLIs.

You probably all like waiting for for that video and to see how how it goes and what it shows and how capy bar

speaks. So let me just uh show the

speaks. So let me just uh show the result to you. So let me roll the video.

It's as you see it's 11 plus minutes. I

will probably just give you like one minute to listen to uh and then yeah everyone can listen to the entire video which is a part of the repository that's

on GitHub hub.

>> Hi I'm Ana Cappy. Today we will talk about safety and security for AI agents.

As AI agents grow in capability, ensuring they operate safely and securely is paramount.

It's crucial they align with your brand values to avoid posing risks to your reputation.

Uncontrolled agents can execute misaligned actions, leak data, or generate inappropriate content. Key

sources of risk include vague instructions, model hallucination, and prompt injections.

I would watch a coding tutorial from Ana the Capivara easily on YouTube.

>> It's so funny. I couldn't stop watching like my brain's like I'm hearing security from an animal. Like you know it's a pattern interrupt like you can't help but watch it. But that was awesome.

>> I'm glad you liked it.

>> Okay, so we've covered so much. This was

such an exciting episode. So, thank you Brandon for coming on and showcasing your demo and having this wonderful conversation with us on such an exciting day.

>> Seriously, thank you Smith. Thank you

Vlad. I've been a little kid in a candy shop playing with all these new models from Gemini anti-gravity. So, seriously,

thank you guys for uh having me on today. And also, I'm going to be

today. And also, I'm going to be checking out that Cappy bar example right after this. So, thank you for sharing that as well, Vlad.

>> Thank you for joining us for today's episode. I hope uh you enjoyed it. Uh,

episode. I hope uh you enjoyed it. Uh,

please join us for the next episode.

Until then, keep your agent smart and are you ready?

[music] >> [music]