I want us to do an agreement today.

When I say something that you like, I

wanted you to answer, "Hell yeah."

>> Hell yeah.

>> All right, let's do a test. One, two,

three.

>> Hell yeah.

You're the best audience. I love

this. I was thinking about what I what I

would say to this audience and I thought

about and even this morning I just

landed like 7 a.m. today in JFK and I

was putting my slides on my way here and

thinking should I go very technical and

like build something and all that should

I go a little more strategic and uh to

Namita's point we actually just pushed

an entire course on like hands-on

building this agent I was like all right

I think the thing that I can share that

they won't find anywhere else is some

insights given the very unique

positioning that we have the crew AI. I

mean, we're sitting on the on the

cutting edge of what is happening right

now and being able to observe what

everyone's doing. So, there's a lot of

interesting insights and tricks and

things that I think could be interesting

for me to share kind of like give you an

insight on what is actually working,

what is noise and things like that. So,

I think that would be fun. Hopefully,

you enjoy it. I want to start with one

number.

That number is 450

million.

It's insane. That's the number of AI

agents we're running a month.

>> All right. There we go. I like it. Thank

you. [laughter]

Uh but I got to say more than the number

on itself, what I'm impressed is the

scale. This is what quarterly executions

of agents and flows are now plotting

using curi attack. Last quarter alone we

executed 1.1 billion agentic executions.

Those are not individual agents. Those

are crews that can have three, four,

five agents. So if you go into the agent

level, it's going to be even more than

that. It's a little bit insane what is

happening. And this number is big, but

at the same point is so small because I

keep thinking about the potential that

this can have. And and we're starting to

see some of the value that actually can

come from this. Now, I assume that

everyone in here knows what is an AI

agent, but I will spend two minutes just

to level set just in case you don't, but

it's going to be two minutes. So, you

all know LLM, right?

>> All right. They're great at creating

content. If you ask them to write an

email for you, they will do so. And if

you ask them to make it funny, they'll

give it a try, right? [laughter] THEY'RE

GIVE IT A TRY. Might not work. But more

interesting is if you ask it to make a

decision. Let's say that you give two

options of emails and you say, "Here's

email A and here is email B. What email

is better?" And these models can

actually pick one say like oh email A is

better because of X Y and Z. And it's

that ability for making reasonable

decisions decisions that fall within

reason that you can explore when you

start to turn these models into agents.

But now instead of choosing emails but

it's just a silly example. They can

actually choose between taking actions

using tools making decisions. So you're

giving it a goal, you're giving it a a

task, and now these agents will decide

how I achieve that. Will I call a tool?

Will I pull a data from a CRM? Will I

push something into my ERP? It's going

to decide how it gets there. So the AI

is what is controlling the application

flow. And that's roughly what an AI

agent is. To put in simple words, it has

agency. and it gets to decide what it's

going to happens next. Now, the thing

that is most interesting about this

because all right, this is very

interesting, but like what is the big

deal about this? Well, it allows you to

do things that you couldn't do before.

One, allows you to create things. You

can create reports, images, videos,

whatever it might be, something that was

not as easy before. Two, it gives you

the ability to react to things in real

time. So you can send for example one

task today and a completely different

task tomorrow and this agent's going to

be able to adapt. It's going to be able

to understand given the data that you're

sending how it's supposed to behave. So

this idea of handling and expecting

behaviors self-healing it's something

very interesting and also this ability

of take actions actually going out there

and make a decision and actually take an

action based on that. So the combination

of these three things actually create a

lot of opportunity for automation and

again automations that you just couldn't

do before.

Now, so what right why should I care

about it? And and this is where I I

things I think are going to get a little

more interesting [clears throat]

on every major technology shift

I have seen on this companies

the the business logic finds a new house

right it goes and moves somewhere. So

when you think back on the 90s, the

desktop apps and there was like those

big players that grabbed that market, a

bunch of the business logic now moved

into that. They're like into those

business apps. Then you go into the

2000s with the web and now Google,

Salesforce, Service Now, they they

grabbed a chunk of the business logic

and a bunch of that moved into that

next. And then you look at the cloud

with all the APIs and that's when we saw

like Snowflake growing like crazy and

data bricks. So all this business logic

now it's moving there as well. So it's

with all this shifts there's always this

ability for this creating not only new

opportunity for new things but also the

part of this logic and this value gets

redistributed and when I look about AI

agents and maybe I'm I'm claiming too

much put our name in there but that's

hey that's what we're going for. uh it's

that I I I see AI agents being more than

an engineering project. It's a

fundamental shift. It is a new layer and

if you want to build a bullish case for

AI agents, it could eat a lot of the

software away. A lot of these systems

goes from having logic in them to

becoming systems of records and this

logic gets moved into this ab

distraction kind of like that we are

calling AI agents now and if you don't

believe me it's okay it's fair you don't

have to but you also see the pattern on

the market I mean come on let's be real

Sergey from Google the guy never even do

interviews and now he's showing up for

podcasts

is taking Bill Gates out of retire

retirement to do live coding with

co-pilot like there there's something

here that the biggest leaders in the

industry are seeing is that it's like

all right this can actually eat software

away this is shaking the industry in a

way that our place is not necessarily

guaranteed we need to get our hands

dirty dirt dirty and actually like jump

into this and figure out what is going

to be our spot so it's very interesting

to see this movement and a little bit on

kind of like how the market's going So

when I think about agents again I I

think in such a broader aspect than as a

as a technology project as something

that will be only living in the realm of

the engineering I think it's going to be

spearheaded by engineers but it's going

to impact entire businesses and we're

seeing some of that if I look at the

biggest companies in the globe this is

how many engineers they have been there

using crew AI what for me is insane to

see like these number of engineers just

kind of like using ki to build those

automations and this again these are not

a small companies these are crazy

companies like this so huge they are so

it's very exciting to see that not only

the main ones that I mentioned but a lot

of the solid business also jumping into

that now

use case time I will stop every now and

then to talk about a use case and I did

heard that hell yeah thank you that one

person in the audience uh I'm going to

sprinkle use cases throughout the thing

right this This one, it's interesting.

It is a CPG. CPG is a company that do

something that you can buy in the

supermarket, right? Something that you

can buy in the supermarket if you want

to think about it that way. They had a

specific process because as every CPG,

if you're selling something to

consumers, you're going to have people

asking for reimbursement requests. And

if you're opening yourself for

reimbursement requests, you're going to

have people trying to fraud you, what is

not fun. So they end up having a very

manual process to verify is this a

fraudulent request or not? And this

involve humans, this involves looking

like is this a new account? Is this

account that has done this before? A lot

of going on in there. But you can

actually automate this with agents. And

so they did the entire process went from

three days to validate this

reimbursement to 10 minutes. What is

insane because now if you're a scammer,

well, I guess like you're unlucky. But

for the people that actually wants the

hand reimbursement, this is such a

better experience. They don't need to be

sitting around waiting three days to get

a money back. Right? So this is just one

use case. We're going to sprinkle more

throughout the presentation and you're

going to see these use case times every

time that we do. But this is a little

bit on kind of what's going on an

example of real use case driving real

value.

Now I want to start digging a little

deeper, right? I want to start talking

about we already got what get what AI

agents are. you already understand kind

of like the bullish vision for AI agents

and what they can be and what a role

they can take into the business in the

future. I want to talk about building

and deploying because I think that's

where a lot of people are spending time

now is on that building phase and

talking a little bit about the ROI of

actually having this things running.

Right now in my perspective there is a

race to the bottom to making building

agents extremely simple

and you're going to have you can see

that because there's many open source

tools there's many paid tools there is

pro code there is low code and

everything in between and everyone's

trying to make this agents easier and

easier but if I plot a line how much

effort it takes.

It's it is funny to me that everyone's

focusing on this [laughter]

because in the end of the day when I go

talk with some of these companies, when

some of these companies reach out to CI

and they're already like trying and

experimenting with agents, they usually

tell me things that hint that they're

stuck in there and they cannot get into

the point phase. basically as they try

to get deploying phase the value the

value of whatever they're using like

diminishes very rapidly and they say we

have all these blockers and we're stuck

here we have hundreds of use cases we

have PC's that are working but I want to

get there and I can't and the reason why

they can't is a mixed bag there's a few

things that are AI related they say like

oh I need I need better evaluations I

need proper guard rails I need all this

and there are many things that are not

even AI related they're like well we

need single sign on we need our back we

need to deploy this on our own

infrastructure more traditional

engineering problems especially when

you're thinking about bigger companies

they're like we cannot get to the other

side of this with any of that and I

think what people don't get is there is

and this is going to sound like a hot

take but you're going to get it there is

no value on building agents

because building is the tax that you're

paying the value is on running them.

Like when you start building a agent,

your ROI is negative. You're spending

time on it. So if I had to plot this

out, so this is the ROI and on the other

side's production.

This is what your ROI looks like. You're

losing money out the way. And then when

you turn it on in production and then

really starts to run many times, then

boom, now it's worth it. But on the

building side, it can be fun, you can

learn a lot, it can be easy, it can be

fast, but you're only cashing out on the

opportunity once they are actually using

it. So I think like there's there's a

gap in there and that's I mean I'm

pitching you career AI. That's basically

what we focus, right? That's the thing.

But you got to see that is a little bit

weird that there's so much demand for

the building and prototyping. Meanwhile,

a lot of the value is actually sitting

on the other side.

So when I look into this, I think about

building and managing and that's kind of

like how we see it. So I think like open

source make extremely easy for people to

build and prototype and then as people

mature and once you get into the value

all those other features it's because

they're actually getting value from it.

So I think it's fair that we can offer a

paid product on that and that's kind of

like how we thought about the company

internally. Now how many of you know

curi?

All right. How many of you don't know

Koreai?

There's a few people that have no clue

what they're doing here. All right,

that's okay. No problem. [laughter] Just

kidding. Uh, in the end of the day, this

is what happens. Like, we we started as

an open source tool, so helping

engineers to build this like

automations. We did a few courses with

Andrew, including the new one that we

just mentioned. And I got to say, it's

amazing like this curses. I I really

appreciate everyone that reached out and

that love it is a lot of effort. If

anything, I don't know if the deep learn

AI team is in there. Thank you for

bearing with me. I'm such a terrible

person to work with. I really appreciate

you bearing with my scheduling and

everything, but we have been putting

these courses together. That has been

exceptionally nice. And then because of

these courses, we realized that people

really want to learn. So, we start doing

certifications. We end up getting a

bunch of people certified for example in

IBM. And IBM became a major partner. Um,

we actually two months ago we closed the

US Department with defense with IBM and

I'm still taking my victory lap because

it was a head-to-head against Palanteer

and we came out on top and hey, there

you go. Thank you. I appreciate it.

[applause] I appreciate it. So, that has

been exciting. But then we start to see

something interesting. jobs being

created like AWS opening a position

specifically hiring for crew AI for

their strategic accounts and uh well we

are working with many strategic accounts

is kind of like the top 25% right the

25% that pays the most we're working

with a lot of those companies so that

makes sense but that was kind of like

building that bullish case for us that w

there's a business in here and that's

kind of like how we start to do a lot of

that and we just kept inviting people to

talk about this what is this idea of AI

agents

Use case time.

>> There you go. I hope it's not the same

person, but I appreciate it. Uh,

financial institution global for global

500.

Main problem. Who here knows what KYC

is? Raise your hands.

[laughter]

I see the pain on you all that know what

it is. Is know your customer is a

process through all companies. For

example, like banks or insuranceances.

Whenever they have a new customer, they

got to do a a big dig on all the data on

that person. They need to understand who

that person is, what they come from,

their taxes, everything in between. Very

manual process. Basically, many

different services that you need to pull

this data from. Then they decide to

automate that KYC process. One, the

first report was more accurate than what

the humans were producing. What was a

little bit low-key scary at first place.

part of the processes went down from one

week to 15 to 30 minutes and the entire

know your customer process actually went

4x faster now using AI agents again

insane to see like the use cases we're

talking about this in such an abstract

level I just want to make sure that I

bring like what are these customers

actually doing and what they're building

so this is another one now the second

thing I want to talk about we talked

about this idea of building and

deploying and where the value is I want

to talk about plan and designing

So when you think about planning and

designing, this is usually a matrix that

I use to think about use cases. They

usually live somewhere in this complex

and precision matrix where you can have

anything in any of these quadrants. An

example of a high complexity, high

precision is filling out IRS forms. It's

complex and you cannot get it wrong. And

an example of like a high complexity but

low precision might be creating

educational material. Like you want this

to be high quality, but there's less

about right and wrong. There's many

different kinds of materials. So it's a

little more easy for you to kind like

get around with that. And a low

complexity, low precision could be help

me prepare a report for a sales call.

What would it be? Probably very simple

to do for an agent. Now we realize

there's people that want to there's two

ends of a spectrum. It's what we're

noticing one use cases that people want

to optimize for agency where they want

to say, "Hey, I really wanted these

agents to figure everything out. What

are the tools they're going to use? What

are like how they're going to work

together? How they're going to delegate

work and all that. In the other side of

the spectrum, we have this idea of a

flow. Like I have a if this then that. I

wanted this to happen first and those

things to happen next." and maybe I had

an LLM in there, but that's all that I

need. So, there's two end of the

spectrums and I think both of those are

fair. I think like a lot of the

engineering principles that brought us

this far still apply and one of them is

KISS, keep it simple, stupid, right?

It's just like, hey, if you don't need

an agent, don't do an agent. So, there's

different side of the spectrum. I think

they complement each other. But there is

a pattern that I'm seeing that is

interesting and I didn't see before. And

I'm calling this pattern agentic systems

because they're not cruise, they're not

agents, but they're also not flows.

They're both intertwined. So you have a

backbone that is deterministic where

you're saying this will happen first,

then this will happen next. But you can

then opt in how much agency you want on

different steps. So maybe I just want

one LM call. Maybe I want to just one

agent. Maybe I want an entire group of

agents that we call a crew. So a good

example would be something like this. I

start the conversation, right? I send a

message. So it's a conversational

experience. That message needs to be

processed like what what I want to do

with that. To do that, I just need an

LLM. I need nothing else other than

that. Then let's say that I just said,

"Hi, like there's nothing to do there."

Well, that should trigger an answer,

goes back, ready for the next message.

But let's say that I have something

super complex that needs to do some sort

of deep research. Well, that could kick

off an entire crew. And now that's when

you want more agents because you want

these agents to research, to query, to

analyze, to create a report, and

eventually bring it back for you to an

answer. So this is a pattern that I'm

finding that is being way more common

nowadays for most of the use cases where

you might want to like some of that

deterministic backbone but you want to

opt in on when you want to bring like

these different levels of agents. So

it's an interesting pattern unclear yet

like is this the future? Is this the

next year? Is this the current state of

things? But it's what we're seeing out

there just getting a lot of traction.

Use case time.

>> There you go. I like it.

Global telecon.

This is a little bit of weird one, so

bear with me. Every use case that we

talked about so far has been efficient

gains, saving money. This is generating

revenue in a way.

And is there anyone that works in a

telecon here?

There's one guy

>> there. What I don't get about telecons

is they never want to do just telecon.

They always want to do other things as

well. Go figure, right? So that's the

thing. This company, they wanted to try

a new line of business. They say, "We

have all this data. We know who these

people are, where they live, what phone

they have, what are their patterns, if

they're prepaid or postpaid. So we have

all this. Can we use this data to define

a credit score and lend them money if

they want to?" And they're like, "Right,

we can. But if they want to do this as a

regular business that means that now you

have to hire entire huge team right to

understand like how you establish all

those things together. And what they

decide to do is build the entire thing

with crews. So they use agents to

basically analyze the customer behavior

analyze the patterns and created this

new line of business. And because

they're automating the whole thing with

agents they can actually get the money

access to money in two days which is

insane. Let me just double check if I'm

being about to be true off stage. I'm

good on time.

>> We have about five if you want then.

>> All right. So, I'm going to fast track

things folks real quick. Bear with me.

Where to end? Needs of the stack. What

is getting in the way? We're talking

about this so we don't need to revisit

it. LMS were adopted to cutting the edge

of these applications. Random people

finding out and doing use cases created

many problems. information leaking use

cases that are not shared. What we're

seeing with agents is more of a central

adoption under a CIO, a CDIO, a CTO, a

COO where they're controlling all the

guard rails, all the different settings,

the LLMs, the PII futuring and then

enabling different teams. So, it's a

different implementation pattern that

we're seeing out there. We have

basically observing this idea of

building integration observing

optimizing, managing and scaling. Those

seems to be the main things that people

care about. What I would translate into

simpler terms, get started building

trust and delivering value. That's it in

the end of the day. So this kind of like

the pattern that we're seeing, you can

explode each of those six things into

many individual features. And that's

where start to get into memory, tools,

knowledge, and all those things. But we

don't need to get there given the time.

But it's where you can kind like explode

those things. And there is the idea of a

stack that is coming together from the

cloud providers to the data providers to

the LLMs and then an entire new layer of

agentic things where basically we are

like we're grabbing that piece and

calling the agents operation side of the

thing. And then there's a little more

that we could talk about the agents

operations platforms. But in the sake of

time I want to tell you about the course

that we just did. If you want to see

more of that, there's a QR code there

that you can sign up for the course. And

I want to go into Q&A. So, thank you so

much. Appreciate it.

[applause]

[cheering]

>> Um, so we have time for a couple of

questions.

>> Give it once.

>> There's one there.

>> Back there. [snorts]

>> [clears throat]

>> Hey, thanks. Uh, great presentation. I'm

curious what you've seen or what you'd

recommend um for like shared context,

persistent context between uh different

agents if is there best practice there

that you that you've seen or um

or

>> when you mean share. So your question is

if there's best practice around sharing

context, you mean sharing context with

these agents.

>> Yeah. So let's say there's like multiple

users using different systems

um and different agents for different

workflows

um for them to have uh some shared

>> yeah context window. It's a it's

actually a problem that we are hearing

from some companies and that is like

well we're using agents but there's many

people building agents in this

organization and we're building the same

things twice. We are building the same

tools twice and there's no like central

thing. What we have been doing is we

specifically have a feature that they

like it's called agent repository. So

all your agents are within that. So you

can reuse them many times over right. So

if you let's say someone in one part of

the team builds something that it's a

very incredible sales agents they can

now reuse that across the entire

company. Same thing for specific tools

is what we have been seeing companies

trying to create this idea of an

internal repository to be reusable

across the entire stack.

>> Okay.

>> Does that answer your question?

>> That was like um that was part of it.

Right. I think I think what we were

talking about is um how to make

different agents and tools and and and

things like that like a the toolkit I

guess agent toolkit [laughter]

>> uh that you share but I I guess just

like the context uh is there like a

constantly evolving

>> oh like long-term memory and that kind

of thing.

>> Well, that's cool. Yeah, we see there's

definitely the idea of different kinds

of memory. There's like short-term

memory, there's long-term memory, and

there's also this idea of entities

remembering things, right? So, as these

agents remember things, how they store

that so as they run over and over, they

don't need to research or do the same

work. So, we're seeing a lot of people

using either rag or graph rag as

backbones to implement memories for

agents. Some frameworks come with them

embedded. we come with something

embedded but also other frameworks also

come with their own implementations but

this idea of memory I think there was a

bunch of implementations early on in the

last two years that have been somewhat

naive I think now there's a lot more

people thinking about like all right

what would be like a very robust

implementation of what this look like

yeah

>> thank you Joe okay one more question

from this side okay and then one last

question after that

>> yeah I I have a use case which we're

trying to address and this is uh to have

multiple agents

uh just carry out a conversation to

simulate a conversation on an ethical

issue

and how would you that work? So just

need some ideas on uh you know building

that kind of an application

if uh crew true AI uh can assist. Yeah,

>> we have one use case that is kind of

like that. It would be kind of like a QA

like I want I want these agents to

simulate something. So it's kind of like

QAing something. Is that is that what

you mean?

>> Yeah, but it's more like you know

it's like a group discussion.

>> Yeah.

>> With multiple

Yeah, we see we see some implementations

of that. Usually what people want is I

want to have one single entry point. So

it feels that I'm talking with one thing

but behind the scenes you might have

many agents that are actually doing the

work. So it feels like it's just a

regular chat GPT but then you ask it to

do some crazy things and it goes and do

what is honestly probably very similar

on the chat GP implementation itself and

you ask it to like do a deep research it

triggers probably a multitude of things

that happens in the background but yeah

that's a very usual implementation we

see that for exploring financial data

for example is a very common use case

>> just one representative to that

>> so with the Jack Gh GPT

they've come up with some kind of an SDK

would be able to develop an extension on

chat GPT to do something like that

>> I haven't seen that on chat GPT no but

you can do something with CI that would

be custom but I don't think chat GPT

allows you to extend that they have the

assistance API that is so much similar

but it's detached from the regular

experience

>> so you need to create your own interface

>> yeah there's a few things that come a

little more package. But yes,