February 28, 2025 • 28 mins
What if your next co-worker was an AI that could think and act like a human?
In this episode of “What’s the BUZZ?,” host Andreas Welsch sits down with Eduardo Ordax, Generative AI Lead at AWS, to explore the groundbreaking world of AI agents and their potential to revolutionize business operations.
As companies race to adopt AI, Eduardo shares valuable insights on how these intelligent agents can evolve from mere task automation to strategic partners capable of planning, self-correcting, and collaborating across functions.
Together, they explore essential topics including:
- The reality behind AI agents that separates them from traditional automation tools
- The importance of starting small to effectively integrate AI agents into existing workflows
- Key technologies and frameworks for building AI agents
- The evolving landscape of AI and the challenges that lie ahead
Whether you’re a business executive seeking innovative solutions, a tech aficionado keeping an eye on the latest trends, or interested in the practical applications of AI in the corporate world, this episode offers a treasure trove of actionable insights.
Ready to unlock the potential of AI agents for your business?
Don't miss this episode—tune in now to find out how you can transform AI hype into real-world results!
Questions or suggestions? Send me a Text Message.
***********
Disclaimer: Views are the participants’ own and do not represent those of any participant’s past, present, or future employers. Participation in this event is independent of any potential business relationship (past, present, or future) between the participants or between their employers.
Level up your AI Leadership game with the AI Leadership Handbook:
https://www.aileadershiphandbook.com
More details:
https://www.intelligence-briefing.com
All episodes:
https://www.intelligence-briefing.com/podcast
Get a weekly thought-provoking post in your inbox:
https://www.intelligence-briefing.com/newsletter
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Andreas Welsch (00:00):
Today, we'll talk about how you can build
your first AI agents and whobetter to talk about it than
someone who's actively workingon that with a lot of customers
in Europe, Eduardo Ordax.
Hey, Eduardo, thank you so muchfor joining.
Eduardo Ordax (00:11):
Thanks.
Thanks so much for inviting metoday.
It's really a pleasure to behere.
Andreas Welsch (00:16):
Hey, Eduardo, I've been seeing a lot of you on
LinkedIn over the last one ortwo years, I see you're picking
up a lot of good momentum, soI'm super excited that you're
spending this time, not justwith me, but with the audience
as well.
But for those of you who don'tknow you, maybe you can
introduce yourself real quickand tell us a little bit about
yourself, who you are and whatyou do.
Eduardo Ordax (00:35):
Yeah, sure.
My name is Eduardo.
I'm living right now in Spainout of Madrid.
I'm working as a Generative AIlead for AWS, managing EMEA,
Europe, Middle East and Africa.
And that means essentially likehow we can help our customers to
leverage all the potential ofgenerative AI as well.
(00:59):
This is what I'm doing.
I'm having so much fun.
I'm having so much work duringthe last few months.
And yeah, that's me.
Andreas Welsch (01:08):
Eduardo, should we play a little game to kick
things off?
Yeah, sure.
Let's go.
Wonderful.
All right.
This one's called In Your OwnWords.
And when I hit the buzzer,you'll see the wheels spinning.
And when they stop you'll see aword.
And I would love for you toanswer with the first thing that
comes to mind and why in yourown words.
And to make it even moreinteresting, you'll only have 60
(01:30):
seconds for you to answer.
Are you ready?
Eduardo Ordax (01:34):
I'll try.
I'll do my best.
Andreas Welsch (01:36):
Good.
So here we go.
If AI were a vehicle, what wouldit be?
60 seconds on the clock.
Eduardo Ordax (01:45):
Wow.
That was, super easy.
I will say it will be aCybertruck.
Andreas Welsch (01:50):
Okay.
Why?
I don't know.
First time I saw it, it was likelike something that it was
coming from the future.
And, many of the things that weare seeing right now related to
artificial intelligence, itseems we are, experiencing some
stuff that it really comes fromthe future, probably two or
(02:10):
three years ago, we could noteven imagine, all the
capabilities, the large languagemodels, but also any other
foundation models.
And I think it's the first timewhere we really believe like all
the things that they're going tocome in the next I don't know,
five to 10 years, they're goingto be being like a more
impactful for all of us.
(02:31):
That's why, I see the Cybertruckand I imagine like a remote
world in the future, maybe onMars.
I don't know.
And I'm thinking of this kind ofartificial intelligence idea,
whatever.
So it's, as simple as that.
It reminds me the cyber truck.
Thank you so much for sharing.
(02:52):
I'm curious what you think inthe audience, what comes to mind
now?
There many ways that we canthink about the, future and, the
futuristic technology.
Some of them are already here,like the cyber truck that is on
the roads in the U.S.
In many ways, I also see theparallels to AI and AI agents.
Certainly it's the big hypetopic of this year.
(03:15):
I think there's still a lot ofmore momentum that this
technology will gain, andvendors that keep pushing the
technology.
But I'm curious, in your workwith customers, what do you see,
what departments are exploringAI agents at the moment?
Eduardo Ordax (03:32):
In my opinion, it's even more than departments
where they are all the customersin general, right?
But even if you are asking melike which kind of customers
they are more willing to startlike exploring and start
implementing agents, these areprobably the ones who have
started as well, in the veryearly days with AI And like all
(03:55):
the customers from very highrelated industries like FSI,
financial service industriesinsurance, healthcare, life
science, because they haveinvest a lot of money, a lot of
time, a lot of resources interms of like also in terms of
like teams and people, they havegained this kind of like
(04:16):
competitive advantage in regardsto work with artificial
intelligence, right?
Because They really know howthey have to structure the data,
their systems, and most of themthey are already moving to the
cloud because at the end of theday if you want to use AI, it's
much better when you're in thecloud, right?
Probably these customers are theones that I'm seeing that it's
(04:39):
been much easier for them tostart implementing agents and
for many different use cases,right?
But most of them the mosttypical use case right now, it's
to automate internal processesbecause at the end of the day
you have a control over thiskind of workflows, right?
Because there is a lot of hypearound agents and sometimes an
(05:02):
agent, it's nothing else than acool workflow, right?
I used to call it like a coolworkflow because most of the
times it's just an LLM who islike making different API calls,
or even doing this kind of likea self reflection.
But it's like that.
Like internal processes, becausethe thing that you can do with
(05:24):
internal processes is you canautomate like things that is
taking a lot of time.
It's taking a lot of like manualsteps.
And you are going to be moreefficient and you are going to
get a lot of, saving on, on, oncosts.
Of course, there are likecustomer facing for like
customer support where you can,I don't know you can implement a
chatbot where you are notexpecting only to provide
(05:45):
information to your endcustomers, but also to take
actions at the end, right?
Hey I want to book a room fornext week in Toronto in a hotel
for four for people.
You are expecting this chatbot,not only to give me information
about the room, but also to makethe call.
to the system, where I can bookthe room, I can give you the
special offers, so I need toconnect through a different
(06:07):
database, I can retrieve thedata, I can send it to you.
But right now, at least what I'mseeing in Europe right now, it's
mainly for internal processeslike, hey, how can I automate
the process to book theholidays?
Or how can I automate theprocess, to request a specific
service?
Or how can I automate To open aninternal ticket, to submit any,
(06:31):
thing, whatever, right?
So this is what I'm seeing themost right now here.
Andreas Welsch (06:35):
Thank you for sharing.
I think that's really goodperspective to also see where
are companies looking at thesethings right now.
And it seems that it's similarto AI and machine learning, RPA,
and those previous hypes, right?
Look for something where, likeyou said, it's your internal
process.
You know how it works.
You ideally have it documented.
(06:57):
There are ways to make it moreefficient, faster, and all the
like.
Now that brings me to my nextquestion that I've been thinking
about.
And that's does it really matterwhat's your first agent that you
build?
And are there specificdepartments?
Is it mainly the IT departmentthat again looks to tease out
(07:17):
the last 20 percent of, Hey, canyou reset my password?
Or I forgot where my nearestprinter is, or I have some other
standard problems.
Is it departments that arelooking at this first?
Or do you see business, finance,
Eduardo Ordax (07:30):
Yeah, I'm seeing a lot of IT departments,
especially we're managing thelife cycle of different
weekends, different claims, andso on, but also like the
business sales marketingdepartments, like how you can
automate.
The relationship with thecustomer, they are like super
interested on this because as Isaid like, when we started to
talk about Generative AI.
(07:50):
Probably the top use case wasabout implementing chatbots,
right?
To automate the relationshipwith the customer, but right now
I'm seeing all these departmentslike, hey, I want to take one
step further, not only toprovide information to my
customers, but ultimately totake actions, right?
But also something that I'mseeing a lot is as part of RAG
(08:11):
use cases, Retrieval AugmentedGeneration.
Like one of the challenges withRAG is as soon as you increment
or as soon as you scale thenumber of documents that you
want to query, the differenttechniques become so
challenging, right?
Because it's not easy to findexactly what you want to query.
So you need to place where tofind the exactly answer to your
(08:31):
query, right?
So we are seeing this kind oflike a gigantic rack where
essentially it's again, the samething is okay, we are taking
different approaches I don'tknow, hybrid search, query
rewriting, query rankerimplemented through agents,
right?
So this is very effectivebecause I can initiate different
journeys.
(08:52):
Just to make sure that I'm goingto retrieve the most important
part of the most important tagfor my previous question.
Agentic RAG is also one of thetop use cases that probably I'm
seeing right now.
Not because of agents, butbecause of RAG that like it's,
probably one of the things thatmost of the customers, they are
implementing.
There is nothing else than aknowledge search, right?
Andreas Welsch (09:13):
I think you mentioned a good point about RAG
documents different approachesto RAG and we see so many
schematics of agents and theircomponents in, our LinkedIn feed
in, newsfeeds day in, day out,right from planning to memory
core modules, everything else.
What are the, key technologycomponents that you actually
(09:35):
need if you want to build anagent or use an agent?
Eduardo Ordax (09:39):
At the end of the day like an agent, it's nothing
else than an LLM doing some kindof action, right?
So if we look into an LLM like,the most simple part, let's say
I'm going to ask something tothe LLM and it's going to give
me an answer, right?
An agent is okay, I'm trying totake one step more.
(10:00):
So I can do this kind ofspecific planning okay, if I'm
asking you to write a document,I'm not going to write a
document in one thought, but I'mgoing to say okay, I need to
analyze what I, need to do, thiskind of cell reflection.
So first of all, I'm going tocome up with different answers.
I'm going to analyze if myanswers that are good or not.
(10:21):
So it's like the model.
It's asking itself okay, howgood, how bad this is going to
be.
And I'm going to make a plan,right?
Okay, I'm going to do firstthis, second this, then this.
But also you can use externaltools, right?
So it's an LLM, as I said okay,you can start a specific action.
So your LLM, because they haveall of these LLMs, most of them,
(10:44):
they have what they call thisfunction calling or tool use.
There is nothing else that, hey,me as an LLM, I can make an API
call to an external system and Ican retrieve this information.
So if I'm asking you for someinformation that I don't know, I
can say okay let's go throughthis website.
I'm going to find the answer andI'm going to amend the answer
(11:06):
that I'm going to provide withthis information.
This is the part of the agentsthat you need is like the LLM,
where you can write this, kindof like a self reflection where
you can analyze.
So it's pretty much like us,right?
If I'm asking you to writesomething, you're not going to
do it in one suit.
Maybe you will write a document,you will do a draft.
(11:27):
And even before a draft, you'regoing to start like planning
okay.
This is the schema of mydocument.
Okay.
This is the schema of mydocument.
I start a first draft and you'regoing to say okay, I'm going to
review my draft.
I will review it and I will sayit's good.
It's bad if it's not goodenough.
Okay.
I'm going to take anotherreview, right?
So these are just LLMs.
It's not like responding at thefirst time, but doing this kind
(11:50):
of unicellular reflection,analysis, planning.
And at the end, even if I needit like I can make different
calls to different tools.
And if we take it in one stepmore, sorry, this is pretty much
like us, right?
Like we are not an experts oneverything, right?
So at the end of the day, wewill have like many agents
(12:11):
working together, right?
So it's not only one agent to doall different steps, but it's
okay, I'm going to put workingdifferent agents on different
tasks.
So you may have an agent that isgoing to be an expert on finding
information in the website.
Okay.
This other agent is going to bethe expert trying to orchestrate
all the others.
Another expert is going to bethe one who is going to book
(12:34):
your travel to Toronto.
So you will have many differentagents, but the thing is okay, I
need even to orchestrate thoseagents, right?
That it's called this metaagents and multi agent
collaboration.
There are like many differentnames, but I will say these are
the four main aspects like selfreflection, planning, tool use
(12:56):
or function calling andcollaboration across agents.
Andreas Welsch (13:00):
Now we've, been talking in the tech industry for
a long time about microservices,have individual capabilities
compartmentalized so we canagain, call them or string them
to together.
That definitely sounds a lotlike that.
And even on a more granularlevel.
But I'm curious too for someonewho might be new to this topic.
(13:21):
If you hear all about agents inthe news, do I have to build all
of this myself?
Do I have to either be adeveloper or do I need a team of
developers?
Are there some things where Ishould put or use things off the
shelf?
What's your recommendationbetween maybe off the shelf and
building this yourself?
When does it make sense to buildit?
Eduardo Ordax (13:40):
Yeah, I think at the end it depends, right?
There are many differentapproaches and there are many
different flavors, right?
If you have something that issuper specific that is Hey, I
want to do this, that is goingto be A, B and C.
You have already like manydifferent agents that they are
going to integrate with yourERP.
(14:00):
Your different systems and youdon't need to create it from
scratch, right?
Hey, you can use the agents fromSalesforce, or you can use the
agents for created virtualassistant, or you can create the
customer support agents.
They're like many differentcompanies that they are creating
these agents.
And then removing all this heavylifting about how to manage the
agents, how to manage theresources, how to query
(14:23):
different databases.
So if you don't have theexpertise, if you don't have the
skills, you Maybe it's better togo through this approach, right?
But they're like all thedifferent flavors, right?
With AWS, you have agents onBedrock.
It's still you can customizewhich kind of agents you are
going to build, right?
You can say, okay, I want tobuild these specific agents for
customer support.
(14:44):
You are going to have access toall these different systems.
But again, it's going to be amanaged service.
It's much easier to implement.
You can give some instructions.
You can build differentworkflows.
So even like the entry barrieris going to be very low, right?
But even let's say you have likevery specific requirements.
You want to create everythingfrom scratch.
(15:04):
It's you have frameworks like Idunno, land graph, right?
So with LangGraph, you cancreate everything from scratch.
Probably the most complicatedpart is like the level of
abstraction is so high and theentry barrier is high as well,
right?
So it's not probably the bestway to go for everyone in the
business unless you have enoughexperience, right?
(15:26):
You have different approachesbased on your skills, based on
your capabilities.
If you build just like thesekind of vertical agents, okay
sales agents or HR agents entrybarrier are very low, but the
capabilities or the likepossibilities to customize these
agents, they're not so high,right?
(15:46):
Because it's like very specific.
You have something in betweenlike using agents on Bedrock,
where you have many things tocustomize.
Essentially, you can customizealmost everything, but still,
you build on top of a frameworkthat it's already defined and
it's there.
And then you can buildeverything from scratch, but you
need to manage your agents, youneed to manage your resources,
(16:06):
you need to do this level ofabstraction.
See you on SOAP.
Again, it depends.
It depends on the customer.
I've seen all different aspectsand I think it's not different
from like a normal software.
Like probably you are going touse SAP, so don't build it from
scratch, but if you need to dosomething very specific, you
will have to do it, right?
So it's, not different from,Traditional software, to be
(16:28):
honest.
Andreas Welsch (16:29):
That makes a lot of sense, and I think that's
good reassurance, right?
Look at what your standardvendors already put out there.
If you need something else ormore than that, or highly
customized to your, not justindustry, but business to build
it, on top of the platform.
Now, I'm curious, in your workhow far advanced do you see
(16:49):
companies be on this journey?
Are they just scratching thesurface and trying to figure out
what are these agents?
Are they any good?
Where can we use them?
Or do you already see, leadersthink about how do we make sure
that they act in, in the sameway?
Maybe that they use the sametone the, same style how they
respond, that they are groundedin, the same documents, maybe?
(17:14):
Not just the information, notjust the business documents, but
a code of conduct or values orsay if they're a finance agent
that they're grounded in IFRSaccounting standards.
How far along is that thinkingthere from what you're seeing?
Eduardo Ordax (17:29):
So I think I've seen a shift over the last few
months, right?
Of course, Europe is differentfrom the U.
S.
In the U.
S.
they are like much more advancedright now, right?
But what I've seen during thelast probably 12 months back.
It was mainly aboutexperimentation, even they were
trying to build something like aPOC that it was going into
(17:51):
production, but it was like,hey, we're implementing agents,
but the scope of these agents,it was super limited, right?
So it was like agents working inproduction, but with a very
limited scope.
Let's say If I want to implementan HR agent to book your
holidays or whatever, it wasexposed only to like a reduced
number of employees, forexample, right?
(18:13):
What I'm seeing right now isthis is starting to change.
So it's not only like to exposethese agents to this kind of
control group, but trying toexpose it to, everyone, right?
So of course here, the mainchallenge is about the outcomes,
it's about the cost.
But right now I'm seeing thecustomers trying to move more
like to this use of agents.
(18:33):
It's in general, right?
It's agents and it's AI at thescale, because what I think is
to be honest, at the end of theday, everything is going to be
around agents.
But because if we think aboutgenerative AI, just about like
LLMs, I think it's useless,right?
Let me explain, right?
It's not like it's useless, butwe as humans, we are expecting
(18:55):
not only generate to provide orto receive information, but at
the end of the day, it's aboutaccomplish a specific task,
right?
And this is what we are gettingfrom by using agents.
That's why I'm saying like LLMs,they are useless because at the
end of the day, all thesecompanies that are expecting to
automate processes, right?
So it's not Hey, I'm expectingto know which is the HR policy,
(19:20):
for vacation.
But ultimately to book myholidays on the, tool through
agents, right?
So what I'm seeing right now,it's more and more customers
that are implementing agents atthe scale.
Of course it depends on theindustry digital native
customers, startups likecustomers that they don't have a
(19:41):
legacy, an IT legacy for them ismuch easier, right?
It's more natural because eventhey don't have the problem
about the data.
They usually have already thedata lake, everything is
integrated and so on.
But also customers from hybridrelated industries FSI,
insurance, healthcare, lifescience.
They are really implementingagents for many different like
(20:01):
use cases, like I don't knowresearch discovery that it could
be like a super complicated.
You can use agents to do that.
So I think within the next 12months, we are going to see an
explosion of your like peopleconsuming these services.
And I think that's good becausethis is the life cycle, or this
is the flywheel that, that wereally need because as soon as
(20:22):
we have more people consumingthese services, we will get much
better models, much more capablemodels, because at the end of
the day, you like to train thesemodels, you need a lot of,
money, right?
So all these companies likeOpenAI and Anthropic.
They need to be profitable.
They need to make a businesswith that.
So as soon as we see more peopleusing these systems at the
(20:43):
scale, we will see a clearimprovement of these LLMs and
foundation models as well.
Andreas Welsch (20:51):
I like how practical you make that and how
you share what you're seeing.
I think it's great to seecompanies experimenting with
this, but also looking at whatare the next steps.
Things that, that we should bedoing and, connecting what you
said with our earlier point ofuse what your out of the box
vendors offer, combine it withsomething that's more customized
(21:11):
on your platform.
Build it there.
I'm just wondering with scalingthe number of agents and even if
there are many very specializedtasks, you put them together in
a collaborative workflow and ina collaborative setting, how do
they, communicate?
How do we make sure that theycommunicate well, that they
exchange the right information,that they expect the right
(21:32):
input, give you back the rightoutput?
We've established, protocols,right?
Things like TCP iP, HTTP, thingslike that in the past that
handled that communication.
Do you think we'll see somethinglike that to ensure
interoperability going forwardas well?
Do we need that?
Eduardo Ordax (21:49):
Yeah you made the 1 million question, right?
Hey, how these models, they aregoing or how these agents, they
are going to communicate.
I still find people that theybelieve that they are like
autonomous, super intelligentsystems.
that they will know how tocommunicate across many others
to accomplish tasks, right?
(22:09):
And we are not there yet.
Like these agents, they are notpeople, right?
But what we can do is we canorganize these agents pretty
much like us, right?
Where we can have differentkinds of organizations.
We can have this kind of likemaster agent that is going to do
like many different things, butwe can have like hierarchical
agents where we can define,okay, this is going to be the
(22:30):
master, right?
And these are going to be theiremployees.
So I can assign different tasksfor different agents.
So I just need to make sure thatI establish a protocol or
relationship between theseagents.
This is much, much easier,right?
Because it's okay, one singleagent connecting through many
different, right?
But it's only like one to oneconnection across all these,
(22:53):
agents and you have control, butthere is only one kind of
connection, right?
Or one kind of protocol, but youmay have as well, like higher
levels of, hierarchies, like youmay have One agent, then you
have different departments andwithin the different
departments, you're going tohave a specialist agent.
So you are including or you areadding one layer more of
(23:14):
complexity, but even you canhave peer to peer agents where
we can work as a team, but thereis not like one single agent
that is going to act as aplanner, but we can work all
together at the same time.
So to do that, there is a mannerto do that, right?
Like even you can implementthat, as I said before, like by
using like a land graph, but youneed to define first this kind
(23:35):
of relationships, and you needto be very clear on the
structures, like every singleagent they are going to
accomplish, and also theparameters and the different
like variables that they aregoing to use between others.
And as I said within AmazonBedrock we recently announced to
reinvent this kind of like multicollaboration agent.
(23:57):
Where you can define theserelationships very easily, where
you can say okay, this is goingto be my like a master agent,
and it's going to communicatewith A, B, and C.
It's going to send all thesedifferent parameters.
It's going to expect all thesedifferent answers.
I'm going to allow them to callthese different systems.
(24:18):
So like we are doing all thesedifferent like a heavy lifting
behind the scenes.
But for you, it's going to bemuch easier, right?
But again, you need to definethe relationships between the
agents.
Otherwise like you can expectsomething that, you won't like
it.
If you just leave them to theirown, that's why, yeah, it's
(24:41):
super important.
And maybe in, in the nearfuture, We will see something
more closer to like autonomousagents where they can, you know,
interact and so on but for sure,we are not there yet.
Andreas Welsch (24:53):
Okay.
Sounds like it's a journey, butmany have already embarked on
it.
Now we're getting close to theend of the show and Eduardo, I
was wondering if you cansummarize the key three
takeaways for our audiencetoday.
Eduardo Ordax (25:05):
I think first of all I will say about trying to
remove all the hype about agentstry to see what is the value,
right?
And the value of an agent, itcould be as simple as, hey, I
just need to build a simpleworkflow.
That is going to do A, B, and C.
That could be an agent.
So don't worry too much aboutall this cap, autonomous agents
(25:29):
and so on.
Let's try to start by thebasics.
My recommendation is to startalso like with tools that
they're going to allow you tomove faster.
Like probably at some point oftime, you will be in a situation
where you can build everythingfrom scratch, right?
But at the beginning it's about,trying, testing, experiment,
fail, and then try it again.
(25:51):
My recommendation is to startwith tools that are going to
reduce this time to market.
And I think like tools likeBedrock or even Vertical Agents
it's going to help you a lot onthe process, right?
Probably what I would say thethird element is to be very
clear on identify thoseprocesses where you are going to
get a lot of value, right?
(26:13):
Like Agents can be superpowerful, But it's not about
overcomplicating the stuff,right?
It's about trying to identifythose processes that right now
seems like super manual, whereyou are losing a lot of money
hey, I need to do this everysingle day, or I'm not even able
to automate it, whatever.
Trying to identify all thesedifferent parts of your
(26:35):
processes that they are lackinga lot of automation.
I'm trying to see how like itwill be by using agents and so
on.
So that will be my threetakeaways.
Try to remove all the hype onagents.
Sometimes it's just pureworkflows.
Second, start by experimenting alot and try to reduce the time
(26:55):
to market.
Use platforms that are going toallow you to move faster.
And third, again, it's allabout, platform development.
Finding the value where you'regoing to implement it.
So try to identify the steps ofyour processes that right now
don't look pretty well and tryto automate it by using agents.
Andreas Welsch (27:13):
Now, that sounds like really, sound advice.
And I know it's, grounded in theexperience that you have, and
that you see every day workingwith customers in Europe.
So Eduardo, thank you so muchfor joining us and for sharing
your experience with us today.
Eduardo Ordax (27:27):
Thanks.
Thanks a lot, Andreas, forinviting me today.
Today, we'll talk about how you can build
your first AI agents and whobetter to talk about it than
someone who's actively workingon that with a lot of customers
in Europe, Eduardo Ordax.
Hey, Eduardo, thank you so muchfor joining.
Eduardo Ordax (00:11):
Thanks.
Thanks so much for inviting metoday.
It's really a pleasure to behere.
Andreas Welsch (00:16):
Hey, Eduardo, I've been seeing a lot of you on
LinkedIn over the last one ortwo years, I see you're picking
up a lot of good momentum, soI'm super excited that you're
spending this time, not justwith me, but with the audience
as well.
But for those of you who don'tknow you, maybe you can
introduce yourself real quickand tell us a little bit about
yourself, who you are and whatyou do.
Eduardo Ordax (00:35):
Yeah, sure.
My name is Eduardo.
I'm living right now in Spainout of Madrid.
I'm working as a Generative AIlead for AWS, managing EMEA,
Europe, Middle East and Africa.
And that means essentially likehow we can help our customers to
leverage all the potential ofgenerative AI as well.
(00:59):
This is what I'm doing.
I'm having so much fun.
I'm having so much work duringthe last few months.
And yeah, that's me.
Andreas Welsch (01:08):
Eduardo, should we play a little game to kick
things off?
Yeah, sure.
Let's go.
Wonderful.
All right.
This one's called In Your OwnWords.
And when I hit the buzzer,you'll see the wheels spinning.
And when they stop you'll see aword.
And I would love for you toanswer with the first thing that
comes to mind and why in yourown words.
And to make it even moreinteresting, you'll only have 60
(01:30):
seconds for you to answer.
Are you ready?
Eduardo Ordax (01:34):
I'll try.
I'll do my best.
Andreas Welsch (01:36):
Good.
So here we go.
If AI were a vehicle, what wouldit be?
60 seconds on the clock.
Eduardo Ordax (01:45):
Wow.
That was, super easy.
I will say it will be aCybertruck.
Andreas Welsch (01:50):
Okay.
Why?
I don't know.
First time I saw it, it was likelike something that it was
coming from the future.
And, many of the things that weare seeing right now related to
artificial intelligence, itseems we are, experiencing some
stuff that it really comes fromthe future, probably two or
(02:10):
three years ago, we could noteven imagine, all the
capabilities, the large languagemodels, but also any other
foundation models.
And I think it's the first timewhere we really believe like all
the things that they're going tocome in the next I don't know,
five to 10 years, they're goingto be being like a more
impactful for all of us.
(02:31):
That's why, I see the Cybertruckand I imagine like a remote
world in the future, maybe onMars.
I don't know.
And I'm thinking of this kind ofartificial intelligence idea,
whatever.
So it's, as simple as that.
It reminds me the cyber truck.
Thank you so much for sharing.
(02:52):
I'm curious what you think inthe audience, what comes to mind
now?
There many ways that we canthink about the, future and, the
futuristic technology.
Some of them are already here,like the cyber truck that is on
the roads in the U.S.
In many ways, I also see theparallels to AI and AI agents.
Certainly it's the big hypetopic of this year.
(03:15):
I think there's still a lot ofmore momentum that this
technology will gain, andvendors that keep pushing the
technology.
But I'm curious, in your workwith customers, what do you see,
what departments are exploringAI agents at the moment?
Eduardo Ordax (03:32):
In my opinion, it's even more than departments
where they are all the customersin general, right?
But even if you are asking melike which kind of customers
they are more willing to startlike exploring and start
implementing agents, these areprobably the ones who have
started as well, in the veryearly days with AI And like all
(03:55):
the customers from very highrelated industries like FSI,
financial service industriesinsurance, healthcare, life
science, because they haveinvest a lot of money, a lot of
time, a lot of resources interms of like also in terms of
like teams and people, they havegained this kind of like
(04:16):
competitive advantage in regardsto work with artificial
intelligence, right?
Because They really know howthey have to structure the data,
their systems, and most of themthey are already moving to the
cloud because at the end of theday if you want to use AI, it's
much better when you're in thecloud, right?
Probably these customers are theones that I'm seeing that it's
(04:39):
been much easier for them tostart implementing agents and
for many different use cases,right?
But most of them the mosttypical use case right now, it's
to automate internal processesbecause at the end of the day
you have a control over thiskind of workflows, right?
Because there is a lot of hypearound agents and sometimes an
(05:02):
agent, it's nothing else than acool workflow, right?
I used to call it like a coolworkflow because most of the
times it's just an LLM who islike making different API calls,
or even doing this kind of likea self reflection.
But it's like that.
Like internal processes, becausethe thing that you can do with
(05:24):
internal processes is you canautomate like things that is
taking a lot of time.
It's taking a lot of like manualsteps.
And you are going to be moreefficient and you are going to
get a lot of, saving on, on, oncosts.
Of course, there are likecustomer facing for like
customer support where you can,I don't know you can implement a
chatbot where you are notexpecting only to provide
(05:45):
information to your endcustomers, but also to take
actions at the end, right?
Hey I want to book a room fornext week in Toronto in a hotel
for four for people.
You are expecting this chatbot,not only to give me information
about the room, but also to makethe call.
to the system, where I can bookthe room, I can give you the
special offers, so I need toconnect through a different
(06:07):
database, I can retrieve thedata, I can send it to you.
But right now, at least what I'mseeing in Europe right now, it's
mainly for internal processeslike, hey, how can I automate
the process to book theholidays?
Or how can I automate theprocess, to request a specific
service?
Or how can I automate To open aninternal ticket, to submit any,
(06:31):
thing, whatever, right?
So this is what I'm seeing themost right now here.
Andreas Welsch (06:35):
Thank you for sharing.
I think that's really goodperspective to also see where
are companies looking at thesethings right now.
And it seems that it's similarto AI and machine learning, RPA,
and those previous hypes, right?
Look for something where, likeyou said, it's your internal
process.
You know how it works.
You ideally have it documented.
(06:57):
There are ways to make it moreefficient, faster, and all the
like.
Now that brings me to my nextquestion that I've been thinking
about.
And that's does it really matterwhat's your first agent that you
build?
And are there specificdepartments?
Is it mainly the IT departmentthat again looks to tease out
(07:17):
the last 20 percent of, Hey, canyou reset my password?
Or I forgot where my nearestprinter is, or I have some other
standard problems.
Is it departments that arelooking at this first?
Or do you see business, finance,
Eduardo Ordax (07:30):
Yeah, I'm seeing a lot of IT departments,
especially we're managing thelife cycle of different
weekends, different claims, andso on, but also like the
business sales marketingdepartments, like how you can
automate.
The relationship with thecustomer, they are like super
interested on this because as Isaid like, when we started to
talk about Generative AI.
(07:50):
Probably the top use case wasabout implementing chatbots,
right?
To automate the relationshipwith the customer, but right now
I'm seeing all these departmentslike, hey, I want to take one
step further, not only toprovide information to my
customers, but ultimately totake actions, right?
But also something that I'mseeing a lot is as part of RAG
(08:11):
use cases, Retrieval AugmentedGeneration.
Like one of the challenges withRAG is as soon as you increment
or as soon as you scale thenumber of documents that you
want to query, the differenttechniques become so
challenging, right?
Because it's not easy to findexactly what you want to query.
So you need to place where tofind the exactly answer to your
(08:31):
query, right?
So we are seeing this kind oflike a gigantic rack where
essentially it's again, the samething is okay, we are taking
different approaches I don'tknow, hybrid search, query
rewriting, query rankerimplemented through agents,
right?
So this is very effectivebecause I can initiate different
journeys.
(08:52):
Just to make sure that I'm goingto retrieve the most important
part of the most important tagfor my previous question.
Agentic RAG is also one of thetop use cases that probably I'm
seeing right now.
Not because of agents, butbecause of RAG that like it's,
probably one of the things thatmost of the customers, they are
implementing.
There is nothing else than aknowledge search, right?
Andreas Welsch (09:13):
I think you mentioned a good point about RAG
documents different approachesto RAG and we see so many
schematics of agents and theircomponents in, our LinkedIn feed
in, newsfeeds day in, day out,right from planning to memory
core modules, everything else.
What are the, key technologycomponents that you actually
(09:35):
need if you want to build anagent or use an agent?
Eduardo Ordax (09:39):
At the end of the day like an agent, it's nothing
else than an LLM doing some kindof action, right?
So if we look into an LLM like,the most simple part, let's say
I'm going to ask something tothe LLM and it's going to give
me an answer, right?
An agent is okay, I'm trying totake one step more.
(10:00):
So I can do this kind ofspecific planning okay, if I'm
asking you to write a document,I'm not going to write a
document in one thought, but I'mgoing to say okay, I need to
analyze what I, need to do, thiskind of cell reflection.
So first of all, I'm going tocome up with different answers.
I'm going to analyze if myanswers that are good or not.
(10:21):
So it's like the model.
It's asking itself okay, howgood, how bad this is going to
be.
And I'm going to make a plan,right?
Okay, I'm going to do firstthis, second this, then this.
But also you can use externaltools, right?
So it's an LLM, as I said okay,you can start a specific action.
So your LLM, because they haveall of these LLMs, most of them,
(10:44):
they have what they call thisfunction calling or tool use.
There is nothing else that, hey,me as an LLM, I can make an API
call to an external system and Ican retrieve this information.
So if I'm asking you for someinformation that I don't know, I
can say okay let's go throughthis website.
I'm going to find the answer andI'm going to amend the answer
(11:06):
that I'm going to provide withthis information.
This is the part of the agentsthat you need is like the LLM,
where you can write this, kindof like a self reflection where
you can analyze.
So it's pretty much like us,right?
If I'm asking you to writesomething, you're not going to
do it in one suit.
Maybe you will write a document,you will do a draft.
(11:27):
And even before a draft, you'regoing to start like planning
okay.
This is the schema of mydocument.
Okay.
This is the schema of mydocument.
I start a first draft and you'regoing to say okay, I'm going to
review my draft.
I will review it and I will sayit's good.
It's bad if it's not goodenough.
Okay.
I'm going to take anotherreview, right?
So these are just LLMs.
It's not like responding at thefirst time, but doing this kind
(11:50):
of unicellular reflection,analysis, planning.
And at the end, even if I needit like I can make different
calls to different tools.
And if we take it in one stepmore, sorry, this is pretty much
like us, right?
Like we are not an experts oneverything, right?
So at the end of the day, wewill have like many agents
(12:11):
working together, right?
So it's not only one agent to doall different steps, but it's
okay, I'm going to put workingdifferent agents on different
tasks.
So you may have an agent that isgoing to be an expert on finding
information in the website.
Okay.
This other agent is going to bethe expert trying to orchestrate
all the others.
Another expert is going to bethe one who is going to book
(12:34):
your travel to Toronto.
So you will have many differentagents, but the thing is okay, I
need even to orchestrate thoseagents, right?
That it's called this metaagents and multi agent
collaboration.
There are like many differentnames, but I will say these are
the four main aspects like selfreflection, planning, tool use
(12:56):
or function calling andcollaboration across agents.
Andreas Welsch (13:00):
Now we've, been talking in the tech industry for
a long time about microservices,have individual capabilities
compartmentalized so we canagain, call them or string them
to together.
That definitely sounds a lotlike that.
And even on a more granularlevel.
But I'm curious too for someonewho might be new to this topic.
(13:21):
If you hear all about agents inthe news, do I have to build all
of this myself?
Do I have to either be adeveloper or do I need a team of
developers?
Are there some things where Ishould put or use things off the
shelf?
What's your recommendationbetween maybe off the shelf and
building this yourself?
When does it make sense to buildit?
Eduardo Ordax (13:40):
Yeah, I think at the end it depends, right?
There are many differentapproaches and there are many
different flavors, right?
If you have something that issuper specific that is Hey, I
want to do this, that is goingto be A, B and C.
You have already like manydifferent agents that they are
going to integrate with yourERP.
(14:00):
Your different systems and youdon't need to create it from
scratch, right?
Hey, you can use the agents fromSalesforce, or you can use the
agents for created virtualassistant, or you can create the
customer support agents.
They're like many differentcompanies that they are creating
these agents.
And then removing all this heavylifting about how to manage the
agents, how to manage theresources, how to query
(14:23):
different databases.
So if you don't have theexpertise, if you don't have the
skills, you Maybe it's better togo through this approach, right?
But they're like all thedifferent flavors, right?
With AWS, you have agents onBedrock.
It's still you can customizewhich kind of agents you are
going to build, right?
You can say, okay, I want tobuild these specific agents for
customer support.
(14:44):
You are going to have access toall these different systems.
But again, it's going to be amanaged service.
It's much easier to implement.
You can give some instructions.
You can build differentworkflows.
So even like the entry barrieris going to be very low, right?
But even let's say you have likevery specific requirements.
You want to create everythingfrom scratch.
(15:04):
It's you have frameworks like Idunno, land graph, right?
So with LangGraph, you cancreate everything from scratch.
Probably the most complicatedpart is like the level of
abstraction is so high and theentry barrier is high as well,
right?
So it's not probably the bestway to go for everyone in the
business unless you have enoughexperience, right?
(15:26):
You have different approachesbased on your skills, based on
your capabilities.
If you build just like thesekind of vertical agents, okay
sales agents or HR agents entrybarrier are very low, but the
capabilities or the likepossibilities to customize these
agents, they're not so high,right?
(15:46):
Because it's like very specific.
You have something in betweenlike using agents on Bedrock,
where you have many things tocustomize.
Essentially, you can customizealmost everything, but still,
you build on top of a frameworkthat it's already defined and
it's there.
And then you can buildeverything from scratch, but you
need to manage your agents, youneed to manage your resources,
(16:06):
you need to do this level ofabstraction.
See you on SOAP.
Again, it depends.
It depends on the customer.
I've seen all different aspectsand I think it's not different
from like a normal software.
Like probably you are going touse SAP, so don't build it from
scratch, but if you need to dosomething very specific, you
will have to do it, right?
So it's, not different from,Traditional software, to be
(16:28):
honest.
Andreas Welsch (16:29):
That makes a lot of sense, and I think that's
good reassurance, right?
Look at what your standardvendors already put out there.
If you need something else ormore than that, or highly
customized to your, not justindustry, but business to build
it, on top of the platform.
Now, I'm curious, in your workhow far advanced do you see
(16:49):
companies be on this journey?
Are they just scratching thesurface and trying to figure out
what are these agents?
Are they any good?
Where can we use them?
Or do you already see, leadersthink about how do we make sure
that they act in, in the sameway?
Maybe that they use the sametone the, same style how they
respond, that they are groundedin, the same documents, maybe?
(17:14):
Not just the information, notjust the business documents, but
a code of conduct or values orsay if they're a finance agent
that they're grounded in IFRSaccounting standards.
How far along is that thinkingthere from what you're seeing?
Eduardo Ordax (17:29):
So I think I've seen a shift over the last few
months, right?
Of course, Europe is differentfrom the U.
S.
In the U.
S.
they are like much more advancedright now, right?
But what I've seen during thelast probably 12 months back.
It was mainly aboutexperimentation, even they were
trying to build something like aPOC that it was going into
(17:51):
production, but it was like,hey, we're implementing agents,
but the scope of these agents,it was super limited, right?
So it was like agents working inproduction, but with a very
limited scope.
Let's say If I want to implementan HR agent to book your
holidays or whatever, it wasexposed only to like a reduced
number of employees, forexample, right?
(18:13):
What I'm seeing right now isthis is starting to change.
So it's not only like to exposethese agents to this kind of
control group, but trying toexpose it to, everyone, right?
So of course here, the mainchallenge is about the outcomes,
it's about the cost.
But right now I'm seeing thecustomers trying to move more
like to this use of agents.
(18:33):
It's in general, right?
It's agents and it's AI at thescale, because what I think is
to be honest, at the end of theday, everything is going to be
around agents.
But because if we think aboutgenerative AI, just about like
LLMs, I think it's useless,right?
Let me explain, right?
It's not like it's useless, butwe as humans, we are expecting
(18:55):
not only generate to provide orto receive information, but at
the end of the day, it's aboutaccomplish a specific task,
right?
And this is what we are gettingfrom by using agents.
That's why I'm saying like LLMs,they are useless because at the
end of the day, all thesecompanies that are expecting to
automate processes, right?
So it's not Hey, I'm expectingto know which is the HR policy,
(19:20):
for vacation.
But ultimately to book myholidays on the, tool through
agents, right?
So what I'm seeing right now,it's more and more customers
that are implementing agents atthe scale.
Of course it depends on theindustry digital native
customers, startups likecustomers that they don't have a
(19:41):
legacy, an IT legacy for them ismuch easier, right?
It's more natural because eventhey don't have the problem
about the data.
They usually have already thedata lake, everything is
integrated and so on.
But also customers from hybridrelated industries FSI,
insurance, healthcare, lifescience.
They are really implementingagents for many different like
(20:01):
use cases, like I don't knowresearch discovery that it could
be like a super complicated.
You can use agents to do that.
So I think within the next 12months, we are going to see an
explosion of your like peopleconsuming these services.
And I think that's good becausethis is the life cycle, or this
is the flywheel that, that wereally need because as soon as
(20:22):
we have more people consumingthese services, we will get much
better models, much more capablemodels, because at the end of
the day, you like to train thesemodels, you need a lot of,
money, right?
So all these companies likeOpenAI and Anthropic.
They need to be profitable.
They need to make a businesswith that.
So as soon as we see more peopleusing these systems at the
(20:43):
scale, we will see a clearimprovement of these LLMs and
foundation models as well.
Andreas Welsch (20:51):
I like how practical you make that and how
you share what you're seeing.
I think it's great to seecompanies experimenting with
this, but also looking at whatare the next steps.
Things that, that we should bedoing and, connecting what you
said with our earlier point ofuse what your out of the box
vendors offer, combine it withsomething that's more customized
(21:11):
on your platform.
Build it there.
I'm just wondering with scalingthe number of agents and even if
there are many very specializedtasks, you put them together in
a collaborative workflow and ina collaborative setting, how do
they, communicate?
How do we make sure that theycommunicate well, that they
exchange the right information,that they expect the right
(21:32):
input, give you back the rightoutput?
We've established, protocols,right?
Things like TCP iP, HTTP, thingslike that in the past that
handled that communication.
Do you think we'll see somethinglike that to ensure
interoperability going forwardas well?
Do we need that?
Eduardo Ordax (21:49):
Yeah you made the 1 million question, right?
Hey, how these models, they aregoing or how these agents, they
are going to communicate.
I still find people that theybelieve that they are like
autonomous, super intelligentsystems.
that they will know how tocommunicate across many others
to accomplish tasks, right?
(22:09):
And we are not there yet.
Like these agents, they are notpeople, right?
But what we can do is we canorganize these agents pretty
much like us, right?
Where we can have differentkinds of organizations.
We can have this kind of likemaster agent that is going to do
like many different things, butwe can have like hierarchical
agents where we can define,okay, this is going to be the
(22:30):
master, right?
And these are going to be theiremployees.
So I can assign different tasksfor different agents.
So I just need to make sure thatI establish a protocol or
relationship between theseagents.
This is much, much easier,right?
Because it's okay, one singleagent connecting through many
different, right?
But it's only like one to oneconnection across all these,
(22:53):
agents and you have control, butthere is only one kind of
connection, right?
Or one kind of protocol, but youmay have as well, like higher
levels of, hierarchies, like youmay have One agent, then you
have different departments andwithin the different
departments, you're going tohave a specialist agent.
So you are including or you areadding one layer more of
(23:14):
complexity, but even you canhave peer to peer agents where
we can work as a team, but thereis not like one single agent
that is going to act as aplanner, but we can work all
together at the same time.
So to do that, there is a mannerto do that, right?
Like even you can implementthat, as I said before, like by
using like a land graph, but youneed to define first this kind
(23:35):
of relationships, and you needto be very clear on the
structures, like every singleagent they are going to
accomplish, and also theparameters and the different
like variables that they aregoing to use between others.
And as I said within AmazonBedrock we recently announced to
reinvent this kind of like multicollaboration agent.
(23:57):
Where you can define theserelationships very easily, where
you can say okay, this is goingto be my like a master agent,
and it's going to communicatewith A, B, and C.
It's going to send all thesedifferent parameters.
It's going to expect all thesedifferent answers.
I'm going to allow them to callthese different systems.
(24:18):
So like we are doing all thesedifferent like a heavy lifting
behind the scenes.
But for you, it's going to bemuch easier, right?
But again, you need to definethe relationships between the
agents.
Otherwise like you can expectsomething that, you won't like
it.
If you just leave them to theirown, that's why, yeah, it's
(24:41):
super important.
And maybe in, in the nearfuture, We will see something
more closer to like autonomousagents where they can, you know,
interact and so on but for sure,we are not there yet.
Andreas Welsch (24:53):
Okay.
Sounds like it's a journey, butmany have already embarked on
it.
Now we're getting close to theend of the show and Eduardo, I
was wondering if you cansummarize the key three
takeaways for our audiencetoday.
Eduardo Ordax (25:05):
I think first of all I will say about trying to
remove all the hype about agentstry to see what is the value,
right?
And the value of an agent, itcould be as simple as, hey, I
just need to build a simpleworkflow.
That is going to do A, B, and C.
That could be an agent.
So don't worry too much aboutall this cap, autonomous agents
(25:29):
and so on.
Let's try to start by thebasics.
My recommendation is to startalso like with tools that
they're going to allow you tomove faster.
Like probably at some point oftime, you will be in a situation
where you can build everythingfrom scratch, right?
But at the beginning it's about,trying, testing, experiment,
fail, and then try it again.
(25:51):
My recommendation is to startwith tools that are going to
reduce this time to market.
And I think like tools likeBedrock or even Vertical Agents
it's going to help you a lot onthe process, right?
Probably what I would say thethird element is to be very
clear on identify thoseprocesses where you are going to
get a lot of value, right?
(26:13):
Like Agents can be superpowerful, But it's not about
overcomplicating the stuff,right?
It's about trying to identifythose processes that right now
seems like super manual, whereyou are losing a lot of money
hey, I need to do this everysingle day, or I'm not even able
to automate it, whatever.
Trying to identify all thesedifferent parts of your
(26:35):
processes that they are lackinga lot of automation.
I'm trying to see how like itwill be by using agents and so
on.
So that will be my threetakeaways.
Try to remove all the hype onagents.
Sometimes it's just pureworkflows.
Second, start by experimenting alot and try to reduce the time
(26:55):
to market.
Use platforms that are going toallow you to move faster.
And third, again, it's allabout, platform development.
Finding the value where you'regoing to implement it.
So try to identify the steps ofyour processes that right now
don't look pretty well and tryto automate it by using agents.
Andreas Welsch (27:13):
Now, that sounds like really, sound advice.
And I know it's, grounded in theexperience that you have, and
that you see every day workingwith customers in Europe.
So Eduardo, thank you so muchfor joining us and for sharing
your experience with us today.
Eduardo Ordax (27:27):
Thanks.
Thanks a lot, Andreas, forinviting me today.
Popular Podcasts
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
24/7 News: The Latest
The latest news in 4 minutes updated every hour, every day.
Therapy Gecko
An unlicensed lizard psychologist travels the universe talking to strangers about absolutely nothing. TO CALL THE GECKO: follow me on https://www.twitch.tv/lyleforever to get a notification for when I am taking calls. I am usually live Mondays, Wednesdays, and Fridays but lately a lot of other times too. I am a gecko.