Who built the most insane AI agent this

week?

>> Whoa, whoa, whoa. It's taking like 5

years.

>> There's a lot that we had to kind of

debug. 403 403.

>> We reset the whole project. I was like,

>> that's when it all went downhill.

>> Who do you guys think should win Google

Cloud's AIHMP?

>> What's it going to be?

>> You're about to find out.

>> My vote is

>> Hey guys, I'm Daniel. I'm here with a

few friends and we've decided to put our

coding skills to the ultimate test.

We're going to tackle a huge problem

every startup runs into. You have a

product but have zero traction. The

answer, a go-to market sidekick for

founders. Oh, and also we're going to do

it in 72 hours. Right away, you could

see our different strategies emerge.

Muhammad and IO figured out a super

clean way to keep context flowing

between their agent. Zach and Lockshi

were basically doing inception building

agents that call a sub agent, which

calls a sequential agent. But me and

Luis, we actually built an ADK agent to

teach us how to even use the agent

development kit in the first place.

Three teams, three totally different

approaches, all starting from the same

powerful framework. And now we're going

to show you exactly how we made it

happen.

Before we could even present our

projects, Abe hits us with a surprise

challenge. Here we go.

>> The goal really of this is to challenge

your agent with real life scenarios that

you see dayto-day. We've laid out six

distinct challenges. Can they

successfully deploy their agent to a

live public endpoint? Can they expose

their tools via the model context

protocol? Can they open for direct

collaboration using the agentto agent

protocol or simulate a thousand

concurrent users hitting their endpoint?

2 hours on the clock starts now.

>> Hello my friend.

>> It is so good to see you.

>> My name is Daniel Eris. I am a third

year at the University of Central

Florida.

>> My name is Luis and I'm a customer

engineer here at Google Cloud and this

is my third AI agent bake off.

>> I thought you had ads on your website.

>> No, no, that's like the actual generated

like

>> Yeah, that's pretty cool.

>> I was like, "Wow, you guys are

monetizing already."

>> We should though.

>> I tackled the multimodal because I

thought, "Oh, that's going to take me 5

minutes." The image is going to input

and then it should append the

transcribed text into the actual

response and then it gets sent to the

research agent that we have. And then

once we have that, it moves on to the

next section which actually has like a

strategy brief which is like a

printable.

>> So like this report outlines the

strategic foundation for ecocon mobile

application and that's exactly what was

on the image beforehand.

>> Wo wo wo. Okay. Just to enable

multimodal. What was the thing that you

had to do?

>> Well, I have like an actual image input

right here and I can just add whatever I

want.

>> You didn't change the model.

>> We were using flash already.

>> Exactly. Yeah. Yeah. Yeah. I also

tackled the A2A because honestly I just

found it interesting.

>> Oh, you got A2A. What you're seeing

there is the model card being displayed.

So you can see you just go to that URL.

>> This is the actual file that I'm

wrapping it with. And then I just hooked

that up with my agent and it pretty much

ported it to a 502 and that was it. That

was the two that I saw that I'm like

okay that's easy. And then now it's like

I don't know 1,000 unique users over a

short period. What about no? But I was

so stressed out for like like I got like

two of the challenges working. The third

one just wasn't working and I was like

nah like it's over.

>> Do you want to do the deploy we were

talking about before? I'm like, "Okay,

what do you think we should do?" And

he's like, "I know how to deploy it."

And I'm like, "All right, cool."

>> ADK has a handy dandy deploy option. And

so you can just tell it cloud run, pass

along a project ID. Boom.

>> You have 25 minutes. So we're just

waiting for this to deploy.

>> Yeah, it's taken like 5 years.

>> Oh no, that's too much.

>> I almost cried when I got the final

thing. Oh, me too. Yeah. No. No.

>> Like I was like,

>> I'm Muhammad. I'm a developer and I lead

a machine learning team at a startup.

>> Hi, I'm I deji. I show developers how to

solve problems in their code using

Google's AI.

>> How do you guys feel for the 808 one?

>> You add a wrapper, you add a few

imports, you can easily create a host

script.

>> That is a value prop of ADK, right? The

80A protocols there.

>> Even though the ADK was just released in

June earlier this year, it's actually

really fully featured and the engineers

behind the scenes have really done a lot

of work to kind of think about what

developers need. I would say it's a lot

more seamless than you would expect.

Which one do you feel most comfortable

with?

>> We can do the MCP. I think

>> the way our system works is a multi-

aent system, right? So, one of them is

specialized for visual identity and it

creates visual assets. There were a

couple of different tools that visual

agent could use, right? I took those

tools, wrap them in an MCP server and

the other server is running

independently and then the agent is able

to communicate with that and then

>> Oh, you guys didn't use fast MCP?

>> Oh, you guys just built it yourself? Oh,

very cool. And then if you guys were to

want to put more of your tools, you

would just do it right here.

>> Exactly. Right.

>> Done.

>> Yeah.

>> MCP, we got that. A2A, we got that. I

think we're in a really good position.

>> Yeah, we did multimodal.

>> Oh, you guys already built that?

>> Yeah. So, if you come over here to our

website superpowers.work.com.

You can log in. So, this is our

onboarding page. We want it to feel like

a real startup. I log in. It's grabbing

my name. IO. Continue. We want people to

feel like they're becoming superheroes

when they use our app. So, essentially,

you can upload a photo. I'm just going

to use my current Google photo. And then

it's going to transform me into a

superhero. And that's me.

>> I'm a superhero.

>> Oh, you just built it now.

>> Yeah. Put the fin put the finer layers

in. Yeah.

>> I'm Zach. I'm a full stack engineer.

>> I'm Lakshmi. I'm a field solutions

architect.

>> As we're going through the challenges,

the server. Okay, that's a bit of a

problem. I'm getting this error when I

try to run it. Why won't anything run?

>> When we got the MCP figured out, all

right, great. We're cooking. We got

this. This our MCP server.

>> We use fast MCP to convert this function

to a tool. So it's basically for

generating images because we do have

something called a social media agent.

It uses nano banana to generate images.

>> Got it.

>> So now with the MCP, you could trigger

that same agent.

>> Yes.

>> Okay. Well, congrats. You finished one.

>> Yes.

>> Took a lot of blood, sweat, and tears to

get here.

>> We got this. We're going to be great.

A2A is basically going to be the same

thing. Lakshmi's got this.

>> So we have a ICP agent,

>> but that's going to be a A2A agent.

Super sleek. Just add this one line.

Just one line to get it ready for agents

to agent to communicate.

>> Yes,

>> I'll rate you off.

>> That's nice.

>> We got this. We can do this.

>> That's when it all went downhill.

>> I think I should be able to do the load

test. I just need to figure out why this

is refusing my connections. I go to the

load balancer and then I start running

into issues once again. Issues that I

shouldn't be having. But

>> you're like, it works but not really.

Okay.

>> It just giving me a 403 every single

time. I ran the script. That worked, but

I just kept getting 403 errors. Just for

whatever reason, the app said it's not

allowed. Sorry.

>> Eight minutes left. It's down to the

wire. I'm sitting there stressing out,

staring at my computer. I'm trying to

make it work. I'm trying to make it

work. I'm trying to make it work. 403.

403. I think that's a a test. No.

>> Okay. I I did say 80% have to pass. Run

a,000 users.

>> Okay.

>> Daniel Luis has done the multimodal, the

A2A, and the deployment. IO Muhammad did

MCP server, the multimmoal input, and

the A2A as well. And last but not least,

Lakshmi and Zack the MCP, A2A, and

multimoto as well. I think we have to do

the countdown.

>> Wo wo wo wo wo.

>> Yep.

>> It's going to take the entire countdown

for my thing to even start.

>> I'll count very slowly.

>> 10. I tried. I tried. Could not get it

to work.

>> Seven seems normal speed.

>> Yeah. A little

>> a little fast. A little fast, I would

say. Whoa. Put your hands down. four

>> 23 three

>> you see so it's um it's deploying to

agent engine you can see you can view

progress logs and we have the agent

engine right

>> open the URL

>> so you guys got four

>> A2A MCP deployment in the multimodal

>> you guys did win regardless you guys got

four so congrats and now we move to the

demo

>> there's a lot that we had to kind of

debug but we actually got it working it

was pretty smooth sailing from there I

will brings a really good expertise in

ADK. I more think about the product

side. So I think there was a really good

match of expertise.

So this challenge is a little different

from the first two episodes.

>> You need to build an agent that is going

to help founders build strategies go to

market. They pretty much gave us like 72

hours to work on it before we got here.

>> The twist was do it virtually and then

you're going to demo it.

>> Obviously this is a unique way of doing

a hackathon. What are some things that

developers should get from this?

>> If you go deep dive into the code,

you'll really understand the different

agent orchestration. When you talk about

multi- aent app, two things that matter

is the orchestration, how these agents

work together and the second thing is

context. How the context is shared

between agents. I think what you can

really take home is the reasoning in

order to play with this agent pattern

>> and you understand what are the

different type of agent that ADK can

build and when you want to put human in

the loop what are the costs how do you

do the tradeoff

>> hello judges

>> hello judges we're presented with this

problem about pretty much generating

anything that could help a startup get

to where it wants to be made this app

called go to market forge it's the forge

for anything that you'll need to go to

the market right you put your input of

whatever idea you have with that We

essentially make everything that

somebody will need including an ad video

of your idea, a PRD, different assets

like a logo and an actual hosted website

as well. So the user input goes in and

then this is put into a deep research

agent which loops around and finds out

pretty much everything from competitors

that have already done the same thing,

what their strategies were, how they

succeeded, different resources that they

might be able to use, like pain points

that they should really try and target.

And so that report will go to an actual

ideation agent which creates a actual

brief. I think something that's looked

over. It's just like the different types

of agents that you can make. Make sure

that you have a legitimate problem that

you're trying to solve and then take the

technology that's available to you,

break down all the complex steps into

smaller uh processes, subprocesses and

then go ahead and implement. We have a

few parallel agents that are all running

at the same time. For example, you have

a limit of 8 seconds with VO and so we

have to like stitch it all together. So

it'll be like a 30-second ad and

creating four or five different videos

takes a pretty long time. And so we

thought we could optimize that by

running them all parallel at the same

time. If you can recognize how we did,

we have all this stuff that takes like

15 minutes to load and we can just have

them running in parallel. It's very

helpful.

>> Yeah. The agent brain sequentially would

probably take about an hour for it to

complete. Oh yeah. So the fact that it

can run in about 15 minutes is itself a

dramatic improvement in terms of

performance.

>> So we talked already about how the deep

research agent uses a looping construct,

right? And it's actually 98% based on

the full stack agent that's in the ADK

samples. And I want to call out uh

whoever developed that. Thank you very

much. That was amazing. So, let's go

ahead and see it in action.

>> To talk a little bit about this design

before we even go forward, these are all

different assets and mockups that have

been created by this platform. I hop on

call with Luis and I show him what I've

done and then he tells me, "You can just

implement the ADK web as your front

end." He told me and I'm like, "Oh, no."

Like, I'm switching this right now.

Like, we reset the whole project. Luis

is going to put in an idea, an Uber for

pet walking and sitting called puppy

prominade. And so, once we submit this,

we have like different stages to this.

They'll all highlight as they're like

happening. we have like an actual asset

server to where all these assets are

going to be uploaded into whatever

session we're in right now where I would

just see it live pretty much as it's

being created.

>> So now I'm going to switch actually to a

more complete version if I may for a

moment. This is the previous run because

it takes about 10 to 15 minutes for the

whole process to run even using parallel

agents. We mentioned that the flow

basically starts with deep research

right and so we generate a deep research

report. This is the report that talks

about the market. Once we do that we can

then create the brief and the brief is

really

>> it's like the backbone I was going to

say. Yeah, it includes an executive

summary. This is used by every sub aent,

right? So, it is stored in the state and

every sub aent is going to use the brief

as its canonical source of truth. And

then from there, the fan out begins and

we start generating all these other sub

assets. Here we generated four mockups.

So, here's the screens for our app,

various different screens in the user

journey. As you can see, they're

properly labeled so that we know what

they mean. These clips can be stitched

together to create a longer clip. We're

going to have our content voice over if

necessary. And so these things are going

to be covered in here. In this case,

we're browsing them one by one, but you

get the basic idea. Also, we mentioned

that we are going to spec out a website.

And so we start with the spec and then

the code gets generated. And so the spec

is right here talking about the various

sections of the page. It creates this

preview for you right here. It

incorporates all the different assets

that we previously generated. Everything

ready to go, including a value

proposition. And in this particular

case, context was a little messed up.

And so we got Rick rrolled by the models

and it actually decided to put in Rick

Ashley which is probably a copyright

violation. So we're maybe going to stop

it right here. So got the basic idea

there and that is effectively GTM Forge

your expert for starting a startup.

Daniel and Luis they had their

application generate a static website. I

loved that.

>> I'm curious how do you handle the result

for the parallel agent? Do you simply

store everything to like Google storage

or like locally? Do you have a

synthesizer agent so that you can manage

the result of PRD and to the mockup or

how do you handle the result for the

parallel agent?

>> First, we make extensive use of state

variables. So the output of one agent is

going to be stored inside a state

variable but on top of that they all

have a shared mechanism for storing the

artifacts into the preview server. The

preview server is using local storage.

>> So you basically store the result for

the parallel agent. the parallel agent

all it does is just kind of say I'm

going to launch all these sub aents in

parallel so it's really the sub agents

that are responsible for saving uh their

their state via callback under most

circumstances the image generator which

is itself a loop right uses after model

call back to save and persist images the

agent that we developed is probably the

most sophisticated one that I've had the

pleasure of working on right super proud

of the work that we've done in order to

produce this really sophisticated set of

agents and sub aents Hey Jes, we are

building superpowers. It's our AI

assisted go to market strategy builder

that will take you through every step of

the journey. The system works in three

simple steps, but these are very crucial

to our agentic system. In the first

step, we have clarification. So we want

to get to know the founders beyond what

they're building. It's like a team

member that you can actually brainstorm

with. After that, we run a comprehensive

analysis. We look at competitors who are

in the market, what exactly they're

doing, what type of strategy they have.

We adopt a datadriven approach. So

rather than relying on an LLM to come up

with something, we actually look at an

AI go to market playbook which we're

going to show later. Then it generates a

very comprehensive report and we

generate assets, images, even fully

automated ads that the founders can use

in a deployable website. So now we're

going to talk about the architecture of

how this is being done.

>> As Muhammad was mentioning, so our phase

zero is when the founder first interacts

with the go to market agent. Essentially

depending on the tone and the you know

the way the uh founder presents the

information the go to market agent may

ask clarifying questions. Oftent times

you want to make sure that you control

how agents operate. So we have an

approval gate using the ADK action

confirmation function to essentially

confirm that the founder wants to move

to the next stage of deep research. The

reason why we're doing that is deep

research is a little bit uh compute

intensive. You always want to make sure

that you're doing the right thing at the

right time. Diving into the research and

analysis we have multiple loops running

in parallel. We have a competitive

analysis phase. We have a monetization

strategy phase, a visual optimization

loop phase, a target local and

demographics loop agent phase, as well

as a thought leader analyzer phase. And

essentially what these loop agents are

doing is in parallel, they're using the

Google search tool to essentially do one

loop, find answers across multiple

questions, and essentially for those

answers that you couldn't find, feed

that back to state, so the next

iteration can then find it and then

build on that answer. And that all feeds

into a sequential pipeline where

essentially we have an outreach channel

optimization agent that consolidates all

the information and then gives back the

consolidated approach for the product

launch. This is a playbook developed by

experts and the idea is rather than

relying on our own prompts we actually

provide this as a knowledge base. If we

go down we can actually see there are

different segments of products and what

has worked before and what hasn't

worked. So we take all this knowledge

and feed that into our market strategy

development agent that is going to come

up with marketing channels for us.

>> So as we know context engineering is

really important. It's not just about

the power of your model. It's how you

present information and with what

context. So once we complete the

research we trigger level two which is

essentially product launch where we

generate a website and we generate

multiple assets. You go from an idea to

a full product launch with the website.

We're now going to walk you through the

onboarding workflow. So what we do here

is we collect a lot of context about the

founder so that when the founder is

interacting with the agent essentially a

lot of that is pre-baked as part of the

context so the agent can provide really

refined details and responses.

>> Here we're actually using Nano Banana to

create a pretty neat looking superhero

avatar for the user.

>> Yeah. So it's taking my Google profile

photo which was retrieved through

Firebase off and like that by using

superpowers I'm a superhero and with

that I will put on my mask because when

you use superpowers you turn into a

superhero and let's go. So we continue.

We're small team. We founded this year.

>> So all of this goes into the context of

the agent when it's doing research.

>> Yeah. All context engineering. We're

based in uh San Francisco with complete

setup. While you're on boarding, we're

actually also triggering a lot of things

behind the scenes using the power of

Google Cloud. So we have this cloud task

Q for example that triggers VO tasks

based off of the image that you upload.

>> So rather than you waiting for a screen

that is loading or task that is being

completed, we wanted to make it very

interactive

>> while your analysis is going. So you can

go get coffee and kind of just take a

break, you know, cuz even superheroes

need breaks. But let's go ahead and just

dive into the interaction. So I'm going

to say something like, I am thinking of

building a virtual reality eyewear.

Immediately, it's going to start asking

me clarifying questions. Gen Z, stylish,

lightweight, fill in the gaps. So it's

filling in a lot of the details for me,

doing a lot of the initial research. I'm

going to say go ahead and kick off.

Remember, we have an action confirmation

gate. So yes, and this is going to go

ahead and kick off the cascade of

multiple parallel agents running

monetization agent, the competitive

analyzer agent. We have the ICP analyzer

agent, the ideal customer profile agent.

We can take a look at the events. So at

this part of the pipeline, we have our

thought leadership analyzer that's

actually kicking off the loop. This

takes some time. It takes 5 to 10

minutes. So we're actually going to go

ahead and show you completed run. And at

the end of it, you're going to get your

first 100 users outreach strategy

executive summary. So we have our target

personas, the priority channels, weekby

goals, and so on. As part of this as

well, you get a bunch of assets. Let's

go over here to products. And we also

converted this to a commercial using

BO3.

>> Since we're using nano banana generate

images, you can actually customize

those, right? So you can bring in your

own logos and stuff like that. This also

deploys the website for you.

>> Superpowers.org/roucts

essentially gives you a really nice

template for your website. So this is

where you'd put some of that ad copy,

put in some of the images that were

generated, but here is giving you the

templates and then you can go ahead and

order. So we hope that we've converted

you to believers in superheroes. They do

exist. And as part of that, we have some

gifts for you. And uh so

>> thank you, man.

>> I can't get the red one.

>> Red one.

>> Wardrobe.

>> We also have superpowers.

>> They didn't get to fully implement it,

but I could see the vision. The little

gift recycling of of you as a superhero

like chilling with your coffee or like

flying through the air. And the idea was

that was going to be the loading screen.

That's a fun way to make the time time

go rather than just like staring at like

a spinner. based on the context that you

are collecting. I would be super curious

and interested to understand like how

that context is helping you to inform

decisions on the loop conditions instead

of having it as a static one, two or

three loops. Rather than it just being a

deterministic loop, we have the agent

and we instruct it to kind of look for

certain things. So essentially gives us

a lot of flexibility to kind of define

when the agent should stop. And the way

we're essentially managing context is

through state injection. The ADK has I

think it's called injection providers

where essentially you can inject state

um directly into the agent context. So

we have prompts for each agent that

we're using like launch page creator. So

you can see that here we are taking

variables from the level one analysis.

So it's taking competitive research,

visual research, ICP research and it's

injecting it in as part of the

instruction for the agent.

>> You talked about context ingestion where

whatever output you're getting you just

inject that output and you talked about

context rot. So how are you managing the

balance between?

>> Okay. Yeah. So there are two components,

right? One is if it was a sequential

agentic loop, right? We would start with

agent one, whatever the results are and

whatever the instructions were, they

were going to go to agent two, right?

And then agent three and four, right? So

basically you're bloating the context.

We're not doing that. What we are doing

is agent one pass on the results and

then agent two basically has that

context rather than whatever the outputs

of the tool calls were. So that's why

while we're like very selective in terms

of what actually gets passed on through

this process I learned a lot about ADK

itself and its capabilities agent

orchestration is very easy especially uh

they're building these new tools into it

having um IO on the team with his

expertise I I got to learn a lot more

>> hello judges we have this great product

that we really believe in but we don't

have any customers and we have no idea

how to get those customers why don't we

just create an agent to do it for us so

we built it. This is Launchpad.

Launchpad is your AI gotom market

strategist for early stage startups. It

takes your product idea. It will do the

product analysis, the market research.

It will create your ideal customer

profile and then it will generate go to

market strategies based off of

everything that it's sort of compiled.

It will give you the user the option of

choosing the right go to market strategy

that is right for you. And then it will

generate assets for you to use. It'll

generate reports. It'll generate images.

It'll generate everything that you need

to use in order to go out there, get

your first 100, get your first 500

customers without having to spend all

that time doing your own research. It

will generate reports. It generates

media images, everything that you need

to know. So, let's say our product idea,

we are really into shoes and we want to

create a new shoe startup called

airsoft. So, it's going to start our

analysis and then it's going to ask some

follow-up questions. So, I have some

stuff here. Let's say our shoes going to

be comfortable. We're targeting

everybody who has, you know, wide feet.

Our key assumptions are people who want

their shoes to fit without having to go

and hunt down that exact shoe. You know,

people who are extremely tall. They got

to find specifically size 15 or 16

shoes. We're going to submit our answer.

And what's happening in the background

right now is our root agent, now that

it's asked our questions, it's gotten

all the information it needs. It's going

to launch this sequential agent. The

sequential agent is running a product

analysis agent. the ICP agent and then

the go to market agent. Each one of

these agents all share this search

agent. This search agent is used by

product analysis ICP or go to market in

order to do the Google research in order

to do the deep research to compile our

go to market strategies. So while that's

happening it goes through each one one

by one and then it's going to save its

output into the session state. So you

can see in the ADK web that this is the

market analysis. We've got the ICP, the

demographics, and then here are the go

to market strategies which are presented

to the user here. So, it's finished its

analysis. It's given us a go to market

strategy that it recommends, and it

weights each option how best it fits our

specific product idea. Let's just go

ahead with the recommended option. So,

now what's happening here is now that

it's given control back to the user,

it's gone back to this root agent here

and we've chosen our go to market

strategy, it's gone to this Launchpad

kit agent. So now it's compiling

everything and creating a single

launchpad kit to help us get started. So

it's going to start with this social

post agent. Basically what it's done is

it's said, "All right, given all of the

things that we know about this startup

idea based on the research and the go to

market strategy. We're going to create a

social media image. We're going to write

a prompt. We're going to send it to the

user, see if they like it. We do. So

we're going to go ahead and improve and

generate." So now it's gone back to the

Launchpad Kit agent and it's generating

this image. Once it's done with that,

it's going to go to our validation plan

agent. And what that is going to do is

it's going to take all of the data it's

compiled and create a one month

validation plan step by step so you know

each week what you need to do to

implement all the stuff that it's given

you. And you can see here that it's

going to step through each one of these

things and give you the process. And

then after that, it'll go to our image

and video generation agent. This is

where it's going to generate any assets

that it deems are necessary. And then

once it's done with that, it's going to

put all of that in one folder, zip it

up, and download it for the user. While

that's is happening, we'll kind of go

through the ADK web and then show you

kind of, you know, what's kind of

happening behind the scenes.

>> It's created like a landing page copy,

an email blast to sort of get your

customers. And you can see that it's

from the validation plan agent, and then

the Launchpad kit agent. Let's check up

here. Download kit is ready. So, let's

go and open that up. And we can see our

product analysis, our ideal customer

profile, our go to market strategy, our

validation plan. And then we have some

images here we can use in our social

media blast to sort of gather our

customers. That is Launchpad. All of the

things that would have taken us weeks,

months, all that money and effort and

time that we' have sunk in, we've done

that in a couple minutes. Now we know

exactly what we need to do to go out and

get our new customers. And that is

Launchpad. Are you ready to invest in

us?

The approach that we took was like build

these small parts. If it has a problem,

we could just cut that branch off and

the whole part would still work. It has

to be as simple as possible. And I think

with ADK and with our approach, we were

able to do that.

>> I get the point of having a search agent

that is uh you know shared amongst all

the others. But I also know that you

know sometimes a little bit confusing

one uh to use an agent as a tool rather

than using sub agents. So can you

clarify a little bit uh why uh you

didn't use this agent as a tool in this

particular scenario?

>> It actually is an agent tool. So the

idea being that each agent retains

context over the Google search that it's

doing. So the product analysis agent

here when or if it actually does need to

do Google research because it doesn't

have enough knowledge on its own for

whatever the founder's product idea is

will use the search agent as an agent

tool so it can retain context and then

save whatever the research is to a

session state variable which is product

analysis and same with the ICP agent and

same with the go to market agent if it

does need to do additional research so

it retains context rather than passing

that context to a search agent

>> and then if we kind of look at it for

the launchpad agent we're using the

whole sequential agent as a agent tool,

right? So, it still has control. So,

there's only one level. You're

transferring the control into the search

agent and then getting it back.

>> This is probably a very common

experience for every engineer who starts

a a new project or framework that you

know they've never used before. I feel

like we don't really uh read

documentation all that well and then we

need to start googling stuff because

something is not working and then the

answer happens to be in the

documentation. read the documentation,

all of it first and then maybe dive in.

>> At least that's what I was thinking.

>> Uh, that's what I've been thinking.

Yeah.

>> Who do you guys think should win Google

Cloud's AIH and payoff?

>> Luis and Daniel.

>> They have the best presentation. They

delivered the message clearly. I can

see, okay, the goal of this hackathon

was to build GTM Forge.

>> The guys were like super excited. They

were super confident about what they

were building, but their generation was

separate. They're generating images and

mockups at the same time. So when you do

that you lose consistency. That was the

biggest drawback.

>> IO and Muhammad

>> I think they spend more time on thinking

how to build a multi- aent system. So I

really appreciate that

>> when we ask their design question like

how do you balance the trade-off their

answer shows that they actually put a

lot of thoughts with the design.

>> Lakshmi and Zach

>> one thing to learn here is we don't

always have to complicate the agentic

pattern. They use a very simple agentic

pattern sequential pattern but even that

gave them very decent very good result.

Sometimes you can solve the problem even

with a single agent with a very simple

multi- agent. You don't have to like

over complicated things to achieve

something that could have been done

easily.

>> GTM Forge versus the superpower team

versus the Launchpad team. Who's your

vote?

>> Gold teams push themselves to use

multi-agent architecture. So that's uh

something that I think we should value.

>> What's it going to be?

>> I will choose team two.

>> I'll go with Daniel.

>> Anie, you're going to be our tiebreaker.

My vote is

>> congratulations everyone for competing

in episode three of Google Cloud's AI

agent bake off. But sadly we can only

have one winner and the winner for

today's challenge is

>> the winner of today's challenge also is

going to get 3,000 Google Cloud credits.

So congrats Muhammad.

>> I think they've thought about every

possible angle. They had a very

comprehensive solution.

>> I think with the tools that we have

right now, if you use them properly, you

can build very complex systems.

>> When you're looking at what we're

building and you see like the end

product, it looks really complex, but

the ADK makes things a lot more

straightforward than you would think.

>> The best way to learn ADK is to look at

the ADK source code.

>> Build it yourself. It's much easier to

understand and figure out what you're

doing if you start from scratch

>> once you actually learn this stuff and

like really get familiar with it. You

could just start having fun with it. It

was quite an experience to say the

least. I got to get some sleep though.

>> My name is Armando spelled with an A

followed by five.

>> Mind you, we get on call first time

talking to him. He's like, "So, is it

Daniel or Daniel?"

>> Just like that. And I'm like,