Latest / Agentic Conversations (formally mlops.community) / The Agent Exchange: Practitioner Insights

Transcript

The Agent Exchange: Practitioner Insights · Agentic Conversations (formally mlops.community)

0:00
Buckle up and strap in. I grabbed three friends to talk
0:03
about two topics related to AI agents in the e-commerce space
0:08
for the great practitioners. Discussion number one, are we
0:13
heading towards the world of generic agents or specialized
0:17
agents? And #2 what is the future of
0:20
companies making shopping agents?
0:28
Let me introduce the three awesome guests.
0:30
I brought in Steven from OLX, who is the head of product there
0:35
working on Gen. AI efforts.
0:38
He's focused on building AI enabled solutions and doing a
0:42
great job, some of which you may have heard of from our last AI
0:47
agents in production virtual conference like OLX Magic.
0:51
I've also brought in a friend of the pod, Kiara, who's back for a
0:55
second appearance. She's a data scientist on the
0:58
Process AI team. And if you haven't listened to
1:01
the episode where we dove into Web agents, we talk about all
1:07
the work she's been doing, testing out different frameworks
1:10
and eventually building out their own web agent framework
1:14
for the Process AI efforts. And last but not least, the most
1:19
opinionated man in AI, Mr. Dimitri, the Senior Director of
1:24
Process AI. This man's a deep thinker.
1:28
For the past few years, I've had the pleasure of being able to
1:32
converse with him about all topics AI, and he's been working
1:36
on Gen. AI products, including them good
1:40
old agents. It's always nice when I get to
1:44
learn from him. As a reminder, we are using a
1:48
simple definition of AI agents, which is AI that can reason,
1:54
execute tasks on a user's behalf, and use tools.
1:58
So a very simple TLDR way of putting it is AI that can break
2:04
out of the chat box. Let's get into our first topic
2:08
of the day. I want to bring in our guests
2:11
and hear their hot takes. So generic versus specialized
2:16
agents. Dimitri, you were introduced as
2:20
the most opinionated man in show business.
2:22
I'll let you kick it off for what you think the future holds
2:26
when it comes to generic or specialized.
2:30
Well, Dimitri, thank you very much for introduction and also
2:33
for giving me, it's a pleasure to be the first person to
2:38
provide the opinion. Well, I think when we talk about
2:42
agent and I like very much the way that you framed the
2:45
discussion. First of all, we need to
2:48
actually look at what type of agents are being built, right?
2:52
Because in the past couple of months we see a lot of companies
2:56
group research teams putting out easier commercial products or
3:01
even open source frameworks to provide functionality for people
3:06
to build agents. And roughly, at least in my
3:11
opinion, we can group them into a number of categories.
3:14
So the first category is what I call generic agents.
3:19
So on extreme case, these are agents that are not specialized
3:26
in terms of reasoning about the task and also not specialized in
3:30
terms of execution of a task. So think about computer use, for
3:35
example, an agent that can take over your computer and just do
3:40
whatever you ask it to do. Think about a web agent that can
3:45
go online, use your browser to either shop for flight, shop for
3:52
an item, or just do some kind of web research.
3:56
So these are agents that most of the time are not specialized.
4:00
And then on another side of the spectrum, you've got agents that
4:04
can do only one task and only on one website or on one service.
4:11
Sometimes they are so specialised that they need
4:15
require a specific API to be able to execute the task.
4:19
So when I look at this, I think that well specialised agents in
4:27
general are the most precise in doing what they required to do.
4:33
But for business wise this is not scalable approach.
4:37
Imagine that you are building a solution that needs to shop
4:41
online. You cannot really build an agent
4:45
or a family of agent that can go and be integrated with every
4:54
shopping site in the world, right.
4:57
So it just doesn't scale on another side of the spectrum, if
5:01
you've got very generic agent. So what we see, they are good at
5:06
everything a little bit, which means that they are good and
5:09
nothing in particular. So they can do a task which
5:13
requires couple of steps of execution.
5:16
The most common that we've seen or we've heard of, they can do
5:21
up to six steps in the sequence before they hit the wall.
5:26
And if your task fall within the short execution pattern, you're
5:31
fine. But if you've got a complex task
5:34
that requires 1020 steps, these generic agents, they cannot just
5:40
achieve a result. So when I look at agent, I'm
5:43
looking at 2 dimensions of implementation, I'm looking at
5:46
reasoning capabilities and I'm looking at execution
5:49
capabilities. And in each domain I'm looking
5:55
for specialisation. So in recent capabilities, is it
5:59
something that is not specific at all?
6:03
Is it something that's somewhere specific?
6:05
So I'm thinking about food ordering agent, right?
6:08
So it's an agent which understand that to order food
6:13
you need to understand preferences of people.
6:15
You actually need to understand people.
6:17
You need to build what is called food ticket.
6:19
And then you need to go online for example and visit couple of
6:23
services providing food to make an order.
6:27
And then on execution part, I'm looking at again agent that can
6:32
just use your computer or use your browsers and being
6:36
absolutely not specialized to an agent that can do well on
6:41
specific websites. So if you look at the latest
6:46
products that we saw, we saw operator from Open AI.
6:51
So in my opinion, this is an example of agent that is fairly
6:56
generic in reasoning part and has a combination of generic
7:02
execution with a specialised execution.
7:07
And my opinion that the future is for agents that are
7:13
reasonably specialized in reasoning and have a combination
7:17
of specialization and execution with the possibility to fall
7:20
back on generic execution pattern.
7:25
Bit of a hot take there at the end when it came to the
7:27
specialization and reasoning, but got to be honest with you
7:31
man, I was hoping for some stronger opinions.
7:35
That's OK though, it's just a warm up session.
7:37
I get it. We've also got Chiara to take us
7:40
through some of this because Chiara, you've been working with
7:43
the generic form of these agents for a while.
7:45
What are your thoughts? Yeah, so actually I started
7:50
working on generic form of web agent, but then quickly deviated
7:56
towards some more specialized agents.
7:58
And at the end, good balance was something in between, like
8:02
Dimitri just said, because the powerful think of agents is that
8:09
there is not a fixed workflow they can adopt based on what's
8:12
in front of them, right? Kind of like a person would.
8:16
And agents can call tools that can be agents themselves.
8:19
So you can have very specialized agents that are still managed by
8:23
a generic agent and vice versa. So one thing could be having a
8:27
specialized agent for shopping that can browse platforms, knows
8:32
how to browse bigger platforms very well, but still can be
8:37
generic web browser when it comes to smaller websites that
8:41
it has never seen. So with agents, it's much easier
8:46
to do to have this kind of mixed approach compared to previously.
8:52
I'm thinking this, this is a discussion generic versus
8:55
specialist that predates agents. It was there, you know, with
8:58
smaller machine learning models versus bigger models, it's
9:03
always been around. But I think agents broke this
9:06
pattern because it's much, much easier to dynamically choose
9:11
what's what's best. And we see it also with recent
9:17
tools that come out like open AO operator, they made a call for
9:22
action for websites to integrate with them.
9:25
So they prefer to use APIs to integrate and do things fast
9:30
while they also can browse the web in in general.
9:35
And I think from point of view of customer interaction, the
9:39
biggest, biggest changes in speed, of course there is
9:44
success rate. It's much better for specialised
9:47
agents if they have their specialised scenario in front of
9:50
them. That's that's a fact.
9:53
But if you have a long enough, long enough time and number of
9:59
calls you need, you can do like with reasoning models, you could
10:02
still get there. The problem is if you're
10:05
interacting with a person in real time, this is not possible.
10:08
So This is why specialised agents are still very relevant.
10:15
If you have something like deep research that well we saw it
10:20
with Gemini Open AI also recently released this.
10:25
Then you can have a generalist web, sorry generalist web
10:28
browsing tool that can be slow. They don't care how much it
10:32
takes, but gets to the answer and the user is happy at the
10:36
end. Speaking of users, Steven, let's
10:41
bring in your product mindset. I know you like to think about
10:45
how users are going to react to these new tools they get.
10:50
Where do you see the world of generalized versus specialized
10:53
going? That's right.
10:54
So yeah, let me bring in the user angle because I think I
10:57
wouldn't be doing my product background justice if I didn't
11:02
look at the bottom line is I don't believe in sort of 0 to
11:07
one full adoption straight away. I think it would be very
11:09
iterative. And I, I think a, a good way to
11:12
think through this, or at least the way I think through this is
11:16
with an analogy of a human assistant.
11:18
So imagine that you get an assistant today and imagine
11:21
they're, you know, just finish uni and they're going to help
11:24
you with whatever. Probably in the start, you will
11:28
just ask them with OK, maybe get me a coffee.
11:30
Hey, maybe you know, here's some some raw notes I wrote.
11:33
Maybe predify it, maybe use chachi peti to do so.
11:38
But essentially, you'll give very small chunks of of tasks to
11:42
them. And then I will trust with
11:44
without knowing their capabilities, I will trust them
11:47
to do that small thing. Well, like getting me that
11:49
coffee. Then over time, as my trust
11:52
builds, I might dare to give them a bit more complex tasks.
11:57
So I might ask them to write a certain report or research a
12:00
certain topic or create a a plan for a certain area.
12:05
And I will change very much. How much context I give to that
12:09
person I will change. And I will also do it only when
12:12
I have a a proven track record of their capabilities and my
12:16
trust in them grows. And I expect this analogy of how
12:20
we would use a human assistant in our lives to actually, you
12:24
know, work very well for how we're going to be adopting
12:27
agents and assistance in our shopping and throughout our
12:30
entire life anyway. So that means, you know, we'll
12:34
start off with a very small specific purchase and we'll see
12:39
how it does there. For example, buy me movie
12:43
tickets to movie X on Saturday night right in this cinema.
12:47
So there's not much to mess up there.
12:49
Essentially, my trust grows and capabilities grow.
12:53
I might do more complex stuff for hey, buy me a new snowboard
12:57
right now. That framework of actual
13:01
abilities versus perceived abilities.
13:03
I think those are the 2 dimensions that are going to be
13:07
needed for people to trust more and more in agents and for
13:10
agents to cover more and more parts of of customer journeys
13:13
and for for users to accept that.
13:17
So the truth today, and this is what Kiara and also Dimitri
13:22
already say today, generic agents are just not good enough
13:26
to really do stuff. There's errors, there's it, it
13:30
will make a bit dumb mistakes here and there.
13:32
So today you need a specialized agent, one that is catered to 11
13:38
needs or or use case or one particular platform and they're
13:43
actually going to be able to do certain things.
13:45
And so I suspect if we follow that framework of, you know,
13:48
actual capability development and then my perception of that
13:51
capability development, we're going to see the first real
13:53
success cases with those specialized agents.
13:55
So customers going to be trying those and adopting those and
13:59
building some confidence with those.
14:01
And so you're going to start to build some trust with that agent
14:04
in that domain. And just because the
14:07
capabilities aren't there, I think specialized is the only
14:10
way today. Now thinking long term, I don't
14:13
think there's a Moat here. I think that capabilities are a
14:18
short term problem and our perception of capabilities are
14:21
also a short term problem. You know, we have O3 now some
14:26
great stuff coming out with DeepSeek.
14:29
There's obviously no sign of stopping here.
14:33
And so where are we several years from now?
14:36
Well, I think several years from now generic agents like computer
14:38
use, like operator, they're going to be fantastic, right.
14:42
And in that world, I don't know if I still need this generic
14:46
agent that I've come to appreciate and, and trust.
14:49
I could imagine, you know, just like you use Google, you know,
14:52
certain Google tools already now it's just common, right?
14:54
Of course you're going to use it.
14:55
And, and I'm sure, certainly Google is going to launch one
14:57
and, and, and more elapsible as well.
14:59
So I would expect that over the longer term, customers going to
15:04
be moving to, to generic ones as their capabilities improve.
15:07
And, and maybe one more analogy here.
15:11
I don't know if you remember how you started using online
15:13
payments and mobile payments, but it feels very similar to me.
15:16
You probably did or like a very small payment somewhere and then
15:20
you started doing it in more places and you did it first on
15:23
those very specific checkout flows.
15:24
There was no generic one, but now you don't care, right?
15:28
It's all just common and you have maybe your PayPal or you
15:32
have your your Google Pay or whatever you use.
15:34
I suspect a similar type of curve there.
15:36
So bring it back to the discussion.
15:37
You know, generics for a specialized, I think today it it
15:41
needs to be specialized because generics not good enough.
15:44
And I think that is also the only way users going to be
15:46
adopting this technology is through these good experiences,
15:50
small and then bigger, bigger, bigger with specialized ones.
15:53
And as the generic ones get better, there will be a move
15:57
there. And then it's just going to be,
15:58
you know, as common as online payments is today with some some
16:01
big names that you use. Okay, I like the idea of
16:05
perceived capabilities and real capabilities, and that delta
16:11
between the two. I think that's one of the
16:12
biggest disconnects that we have right now in the AI world.
16:17
We see so many different demos around the Internet and it makes
16:21
us say, whoa, look at that, what can AI do?
16:25
It's incredible. And then we go and try and play
16:28
with it and it falls flat on its face.
16:31
I actually think that Steve made a very important point, right?
16:35
And especially what I like about his argument is this historical
16:39
perspective. I'm coming from slightly
16:41
different technical background. So many many years ago I was
16:44
working in IoT space, right? And we also had this discussion,
16:49
OK, what the future will be in IoT in terms of network
16:52
protocols, you remember those? Would it be IP like for the rest
16:56
of Internet or would it be specialised protocols?
16:59
And everyone in the field was saying, well, you cannot do full
17:03
IP stack on those small devices and you need to be really
17:07
specialized. You need to, you know, be real
17:10
time. You need to be light.
17:11
You need to be this and this and this because otherwise it
17:14
doesn't work. What people forget about is what
17:18
Steven actually suggests to think about temporal dimension,
17:21
right? So yes, we cannot do many things
17:24
now, but probably as the most wide, widely adopted, most
17:31
generic solution will win because yes, LLM will get
17:37
better. So it will better get better in
17:40
reasoning. So yes, it will be able to say,
17:44
OK, I've got I can cover food ordering, I can cover buying
17:50
items, I can cover buying trips or whatever.
17:54
So it will be better in reasoning over time and I think
17:58
it will be better in instructing execution part of or execution
18:05
agent or execution part of an agent to go online and do
18:09
something. So as Reese, I agree very high
18:12
likelihood that in the future will just go for generic, but
18:17
for some time it will be indeed very much specialized agents.
18:24
Another aspect that I want to probably mention is that part of
18:29
this discussion is actually not technical.
18:32
You know where technical people, Steven, sorry, I bring you into
18:36
this group of technical people as well.
18:38
But as a technical people, we have a bit of myopic view on
18:43
technology because we say, OK, this is what is possible, this
18:46
is what is reasonable. What I'm also looking at when
18:52
talking about agents in e-commerce, what is the dynamic
18:55
between companies making agents and companies making websites or
19:01
platforms, right. So you can say, OK, generic
19:05
agent will be able to do all the things on the platform, but most
19:12
likely it will not be able to do it effectively, right?
19:16
So think about it's agent which will mimic how we behave on the
19:19
platform. I don't know what's the last
19:21
time you did shopping? I'm absolutely sure I did not
19:25
use an optimal path on Amazon or on the booking or any other
19:31
website, right? So I clicked around like a
19:34
monkey. I look at things that probably
19:37
were not relevant to the end result.
19:41
And now imagine agents doing exactly that.
19:45
So specialization will be still valuable for platforms.
19:51
Let's take something into account though.
19:54
Companies make a lot of money on those random clicks around the
19:59
website because I myself have gone to a website countless
20:04
times with the intention of buying only one thing and being
20:08
enticed into buying a few extra items because of.
20:11
Those damn recommender systems that know too much about me and
20:15
my vices. So how do you cope with that in
20:18
a world where agents just go and get you the thing and they in
20:23
theory are not susceptible to recommender systems?
20:28
I did think about that and the answer is that it's hard look,
20:33
so the ideal agent can execute a let's say a buying agent, right
20:40
shopping agent can help you outline the right criteria for
20:43
what you're looking to buy and then find every possible match,
20:47
right. So recommenders are irrelevant.
20:49
Ideally, this agent can go through the entire inventory on
20:53
a certain platform and just find that perfect one And and
20:56
actually not just one platform could be multiple right.
20:58
If we if we talk about generic ones.
21:00
So this basically means that the the better access that this
21:07
agent has to the entire inventory, the better job they
21:10
can do finding and retrieving those things.
21:13
So it's probably not possible for that browsing agent to just
21:20
go click page by page by page throughout all these millions of
21:23
listings in some cases. So there will need to be some
21:27
kind of other, yeah, like a middle layer here.
21:31
That could mean if it's an external or even a bit, let's
21:36
say a bit of a, a competitor or, or a threatening way, you could
21:40
see it. Like maybe they would have to
21:41
scrape all your inventory all the time, have that in a
21:44
database and then make that database searchable.
21:46
So basically scrape inventory, then create some kind of
21:49
quadrant, set some quadrant database and have the the the
21:54
agent look through there. But another way could be that
21:58
platform actually opens that. So you open your doors for
22:01
agents and you make it easy for agents to find the right thing.
22:04
So recommender systems, they become different.
22:06
They're not showing you what is recommended, but rather gives
22:10
you access to the, you know, all the embeddings that have been
22:14
calculated in some stored in some vector database.
22:16
And you just say, OK, here it is, look what you need.
22:20
So yeah, don't have a strict answer on where things will go,
22:24
but just some thoughts. If I may add to that, I think
22:27
platforms could even use their own recommended model for that.
22:32
So they could decide to expose an API where they the results
22:36
that they propose are already ranked by their own engine.
22:40
And I believe that could even be a solution for ads in that
22:44
regard because they could add content that is boosted in that
22:50
actual response that the agent could use and then present to
22:53
the customer. So I think there are a lot of of
22:56
ways here and it will be really important to see how the user
23:03
interacts with with all these new modalities.
23:07
And to add add to the previous point, there is this kind of
23:15
conflict between the generalist agent and specialized agent.
23:21
But if specialized agents open new use cases and we see
23:24
adoption through through the those use cases, this might be
23:27
that this generally this specialized agents become the
23:31
general generalist. And on the other hand,
23:33
generalist agents can open the door for specialized agents and
23:38
adding new modalities, new tools and so on.
23:41
So yeah, I think customers at the centre of all this and end
23:46
users would be really important. Yeah.
23:50
OK. I want to highlight two things
23:52
you said which I have been thinking about at length.
23:58
One is the recommendations that come back to you and the answers
24:02
that are fed to the user. I can see that being something
24:05
that you get when your task is completed.
24:08
And then there's another part that says, do you want me to do
24:11
any of these other tasks? That is becoming a fairly common
24:15
UX design pattern we're seeing with agents these days.
24:18
So it's not that big of a jump to think that that's going to be
24:23
next. Then the second part that you
24:26
said, Chiara, which I had not thought about, is the
24:28
generalized agent being general at the beginning of the flow and
24:34
then it kicks it off onto a specialized agent to accomplish
24:38
the task. It doesn't feel like that far of
24:41
a stretch because there are some design patterns with agents that
24:44
that's happening in. So once maybe we build out a
24:48
strong network of specialized agents, the generalized agent is
24:52
just there to do air traffic control in a way.
24:56
I think we'll enter a world where you want to build tools
25:00
that agents can use, right? So you want to, and in that way,
25:04
you could think of a specialized agent internally built on a
25:09
certain platform, just offering its capabilities in a sort of
25:13
API to another agent, a generic one like let's say Google's one
25:17
or whoever will build it. And then you get a bit of best
25:20
of both worlds thing. But yeah, I think the problem
25:27
there is not so much on capabilities, but it's rather a
25:29
business strategy kind of thing. Do you want that?
25:33
Right. And I think that's a good bridge
25:36
to the next topic as well. To add to what Steven said,
25:41
we've been experimenting a lot now with LLMS being used as
25:46
recommender systems. So an LLM that knows the user,
25:50
knows what they've done and can put rationale inside all this
25:54
behavior can actually provide input to another LLM during the
26:00
recommendation. So there are already agents that
26:04
work together with other agents to provide a better experience
26:08
for both user and platform. Since we are already kind of
26:13
drifting into this space, we might as well talk about the
26:17
second part of this discussion, which is around the questions of
26:22
what is the future of companies making shopping agents.
26:27
And Steven, you brought in a few strong points, so I'm going to
26:31
let you kick it off because really there are so many
26:34
different parts to this. Where do you want to start
26:37
first? I'm probably not the best to
26:42
understand or or or explain what those companies building those
26:46
agents will do, but I think I am well positioned to answer what
26:50
existing e-commerce platforms will do in response.
26:53
And this will guide then, of course, what those first
26:56
companies, what they'll be doing.
26:58
So I suspect that e-commerce platforms and I come from the
27:02
classifieds marketplace world, so that's used goods, cars, also
27:06
real estate jobs. I believe that these platforms
27:10
will start being very defensive in the in the beginning and then
27:15
we'll move to embrace it. So what do I mean, the
27:19
defensiveness would come from the fact that these agents,
27:23
especially if it's a, you know, a third party shopping agent
27:26
browsing your platform, it removes your customer
27:30
connection, right? So it damages your brand
27:32
loyalty. And these brands have many years
27:35
of investment behind them, hundreds of millions in some
27:37
cases. It also makes it hard for these
27:41
marketplaces to get clear customer signals of what's
27:44
happening on their platform, right?
27:45
As agents start browsing, you need to find some way to to
27:49
identify those and kill that noise.
27:51
Otherwise, you know what, it will be very hard to do any
27:54
customer research or build for real customers.
27:58
And, and I think a, a small analogy here is that of price
28:02
comparison sites. I think we've seen that
28:06
typically classifieds websites don't like to cooperate with
28:09
those. So with those aggregator
28:11
integrator type type players, because again, you, you
28:15
disintermediate, right, you lose the connection to that to that
28:17
customer. So This is why I expect
28:21
platforms to start being defensive and I think they will
28:24
take defensive actions to make the platform harder to use for
28:26
agents, not easier. So what what are those?
28:29
Well, think of the type of things we're doing against
28:32
scrapers, right? So you have both detection that
28:36
you will look at, of course, at the IP addresses where things
28:39
come from, only allowing for example, traffic from
28:41
residential IPS in our data centers.
28:44
In anti scraping practices. You have things like honey pots,
28:47
which are essentially, you know, once you identify a scraper, you
28:51
don't block them, but you in fact sent them to a piece of
28:54
your database or to a piece of your site that a normal user
28:57
would never see. You just start generating random
29:00
data just to overload them and poison essentially the data that
29:03
they're scraping. They wouldn't know any more
29:04
what's good and not these kinds of tactics, and I'm sure there's
29:08
many more will probably be start to be deployed by platforms who
29:13
don't want agents on their on their terrain.
29:17
And I don't think that will work so well.
29:20
So now my, my story changes. I think that they will try and I
29:24
think maybe for a while that's how it will be.
29:26
But then at some point we'll see that agents are here to stay.
29:30
You know, they will, they will permeate society lots and lots
29:33
and lots of users. So then it becomes a kind of,
29:36
you can't beat them, join them. And then I actually expect,
29:40
yeah, we'll, we'll try to make it easier for other agents to,
29:43
to be on our platform. So that are those kind of things
29:45
I just mentioned earlier. So having some, some, some APIs
29:49
open or allowing them to, to look through our vector store
29:53
or, you know, catalogue stuff or whatever, there's, there's
29:57
multiple ways we could help agents, give them tools to be
30:00
more effective on our platform. You can already see that this is
30:03
what Open AI intends to do if they're operator, right?
30:05
They want to have partnerships where they can tune towards
30:09
those sites. And I think, yeah, trying to
30:12
become a partner and trying to help a tool like operator be
30:15
more efficient is something that platforms will start doing.
30:19
And this will really mean a different monetization model for
30:22
classifieds or or marketplaces. It means that the traditional
30:26
ways of just selling eyeballs and and views, that's gone.
30:30
So we need to have other ways and there are ways, I'm certain,
30:35
but it will really change from quite fundamentally the business
30:37
model over time. So yeah, I think we'll start off
30:42
defensive and I think that we'll have some success early, but
30:47
then fail and then I think we're going to change.
30:50
We and I, when I say we, I mean every single marketplace in the
30:53
world is going to open up for this because if they don't, they
30:58
will lose to someone who will an end state.
31:01
You could think about, you know, a marketplace brings together
31:04
supply and demand. So if you don't open up, you
31:06
know, I could just imagine some third party even like even
31:10
decentralized, right, some some blockchain database where all
31:13
offer sets and we're all demands can just look for that.
31:16
That's possible. So you you better open up and
31:19
make it easy because that would, you know, give you a right to
31:21
play. Well, Dimitri, I know that we
31:25
have talked at length about whether or not you expose your
31:28
data to the Internet or you keep it for yourself to create a very
31:35
specialized agent for your platform.
31:37
What are your opinions here? Indeed, I think it's a very
31:40
interesting question. And essentially, if you look at
31:45
what is happening with agents nowadays, and this is also
31:49
somewhat related to our discussion about generic versus
31:52
specialized, right? Because generic agent in a way
31:59
is a stick in the hands of a company making them.
32:05
And the carrot is we can make more specialized agent and we
32:09
can have an agreement with a platform, right?
32:12
So the platform has a choice at the moment, which is either
32:16
collaborate with agents or acknowledge that there will be
32:22
this generic agents that are indistinguishable from normal
32:26
users coming to platform and shopping on your platform
32:31
anyway. And I agree with Steven, in my
32:33
experience as well, when I talk to different platforms, the
32:37
immediate reaction is we're going to block it.
32:39
But this is this reaction lasts for like 30 seconds because it's
32:43
a cat and mouse game and people very quickly understand that.
32:48
Well, you start blocking it, you make life for actual users so
32:52
much difficult that it becomes non practical.
32:55
Now, one of the analogies I use when I think about this is you
33:01
remember the story of Spotify and studios, music studios,
33:06
right? So you basically have artist,
33:09
right? And then you've got labels.
33:12
I call them Studio. I probably should have called
33:14
them labels. And the labels, usually their
33:18
main focus is to find the talent and to basically subscribe them
33:25
so they become part of a label. So if they produce anything of
33:28
value, it's to be sold by labels.
33:31
And now Spotify comes and Spotify says, look, we're gonna
33:36
let people listen to what they want and we want to go and
33:40
completely ignore the way that you package stuff.
33:43
So we're basically not gonna kind of care about your
33:48
interest. And to me, it sounds very
33:50
similar to what is happening with shopping Age and the
33:52
platforms at the moment, because the risks are all the same,
33:56
right? It's what you mentioned, the
34:00
view of users disappears. So labels do not know what
34:04
people like to listen to. And as a result, they cannot do
34:09
their job well. And also the way that risk
34:13
profile presented is very similar because Spotify was, OK,
34:19
we're going to go around you labels and source eight source
34:26
musicians. And now with AI agent, shopping
34:32
agent, it's very similar because risk that we see is you've got
34:37
this agent which goes around the platform and just source goods,
34:43
services from suppliers directly.
34:47
But what happened eventually labels and Spotify struck a deal
34:51
where Spotify says, look, we're going to focus on serving user
34:55
needs and you're going to focus on making sure that the pipeline
35:01
of artist does not dry out because this is what you're good
35:05
at. And to actually have mutually
35:10
beneficial relationship, we're going to share with you usage
35:15
data so you can make your decisions better.
35:19
And because now we're going to use all your catalog, we'll also
35:23
give you a share in our company. So if we grow, you grow.
35:28
So they created this codependent relationship, which also allows
35:33
them to focus each of their on their strengths.
35:36
So if I look at what is happening now and especially
35:40
with open AI operator, I see exactly the same dynamic because
35:47
you've got exactly same carrot and stick open the ISS.
35:50
Look, we can use generic web agent on your platform.
35:58
So we're still going to play well, we're going to buy some,
36:02
allow user, buy something from you, but we better focus on user
36:07
experience. So let's work together.
36:10
If you allow our agent on platform, if you do not block
36:14
us, if you present a optimized version of your website or even
36:20
give us API access, we give you user data so you do not lose it
36:25
to your point, right? So you keep to know who your
36:28
users are and then you can start focusing more of your energy on
36:34
getting suppliers in, on building this domain
36:38
intelligence in and allowing us to actually just focus on
36:44
customizing, personalizing experience of users.
36:49
So in a way for platforms, it could be a blessing, right?
36:52
Because on the platform side, you actually now do two things.
36:56
You, you basically need to deal with your sellers or iPhone with
37:00
a restaurant or take a lot with suppliers.
37:03
And at the same time, you put a lot of effort in fostering this
37:07
user relationship in the building user interfaces.
37:12
So a agent says, OK, we're going to take this app apart from you,
37:17
but in exchange, you're still going to get user data.
37:22
So this was my analogy. I think I could have presented
37:26
it somewhat better, to be honest.
37:28
But this is the thing. And actually looking at our
37:32
companies and I think that we discussed it back and forth
37:34
already a couple of times, I think that yes, we're going to
37:40
prevent agents from going to our platform.
37:43
It's like the first reaction, you know, the second most
37:46
typical reaction. They say we're going to build
37:49
our own shopping agent. And as a technical person, I'm
37:53
like, yes, let's do process e-commerce agent because this
37:57
would be fantastic opportunity for us to apply our technical
38:02
expertise to have fun and so on. But from business perspective,
38:07
you just need to ask, OK, so your platform, which game you
38:12
want to play? And if you built your own
38:16
e-commerce agent, for example, how do you maintain one of the
38:20
core value propositions to users, which is multi platform,
38:26
right? The idea is if I go to agent, I
38:29
say I want to buy this bicycle in Poland.
38:32
I need to be able to go to OLX Poland, I need to be able to go
38:36
to Allegra who is competitor. I need to be able to go to what
38:41
Amazon in case I've got like a well priced bicycle and so on.
38:46
So if it's my agent, I'm basically to have a value
38:52
proposition and open it to competitor businesses, right?
38:56
Huh. I like this idea of the
38:59
distinction between sourcing and shopping agents in B to C versus
39:03
B to B1. Thing that you make very clear
39:07
here is that you're basically creating a new platform that is
39:13
almost an abstraction of many different shopping platforms.
39:17
And if you're not careful, you can quickly lose scope of how
39:20
difficult that is by just thinking, oh, here we're going
39:24
to create an agent that can go and do everything that 10
39:29
different platforms can do. But let's keep this rolling.
39:32
Kiara, I know you have some strong thoughts, so I want to
39:35
hear what you have to say about this.
39:38
Yeah. So actually my opinions are very
39:40
similar to Steven and Dimitri on these that and now we're in a
39:46
stage with platforms are trying to block them because it's a big
39:49
change. I feel like platforms interests
39:54
and user interests don't always align.
39:55
So as a user, I want to find the best deal possible.
40:00
A platform may want to present certain items that maybe are not
40:04
the best deal for the user, but are the best deal for them.
40:10
So. Yeah, I think this is a bit
40:13
scary in general for for platforms.
40:17
But as the user, we'll expect more personalization, we'll
40:21
expect an agent to help them discover products.
40:25
And I think they will kind of expect that the agent will give
40:30
them the best deal possible in order to trust this agent.
40:34
We're not yet there. Users don't trust agent.
40:36
We're still in the first era of online payments to yeah, to give
40:43
the analogy that Steven gave before.
40:45
But as more trust will be put into the edge and then more
40:50
expectations will come from customers and I think platform
40:55
will start to give in on this. There are also more use cases
41:00
that could be enabled. We've been thinking a lot of
41:03
these at these because it is a bit of doom scenario, like it's
41:09
very easy to say all the potential drawbacks that agents
41:13
could have for platforms. So they're clear, but there are
41:16
also new opportunities. For instance, they could
41:19
automatically integrate with smaller vendors and present the
41:22
products directly and have an automated pipeline to integrate
41:28
with their own payment system. So user could buy and search for
41:32
items on the platform itself using an agent and then purchase
41:37
the product from another platform.
41:38
So it will be much easier to work as aggregators here.
41:44
And the other is sharing customer data.
41:48
I think in future recommended systems, it will be more
41:51
important to give rationale to really understand the user, not
41:55
just having a fine-tuned deep learning model and on user and
42:02
items embeddings, but actually having something that can give
42:07
an explainable answer. So we recommended this product
42:11
because you've been browsing this and that, because you like
42:14
this type of items, because next month you're going to run a
42:16
marathon and I need to give you the best shoes.
42:19
So all these will enable a lot more interactions with customer
42:24
and customers are going to expect them.
42:27
So there is one thing that you mentioned that I think we did
42:31
not cover at all. And I also to be honest do not
42:35
see being covered in general, which is selling agents, right.
42:41
So we basically say, OK, shopping agent.
42:43
So I'm as a user, I would like to go and I want to purchase
42:47
something, but there is so much need also for selling agent.
42:52
It could be I'm as a user have something I don't need, I want
42:56
to sell or I'm as a manufacturer would like to offer this product
43:00
or I'm as a service provider have this to offer.
43:03
And my guess, and it's a pure speculation and we're not
43:07
talking about selling agent at the moment because there is no
43:12
infrastructure, if you will, to connect selling agent with a
43:17
shopping agent, right. So this is a part which is
43:19
missing. But yeah, at the moment somebody
43:24
comes up with this infrastructure, then there will
43:28
be a gate opening to building selling agent and then the
43:34
relationship between selling shopping and the platform in the
43:40
middle will change once again and we'll have another
43:42
discussion trying to predict the future.
43:46
Yes, trust is so important and I want to 2nd that.
43:50
I have always thought about trust from the perspective of if
43:55
I'm using an agent and I give it my credit card number, am I
43:58
going to have money in my bank account tomorrow or is it just
44:01
going to drain it because it doesn't understand the
44:04
requirements well enough. But here you're talking about
44:08
how quickly we can lose faith in agents if they're not giving us
44:11
the best deals. You can see a world where if a
44:15
buyer is being shown answers that someone paid to get in
44:20
front of them, the buyer will quickly revert to doing its own
44:24
research because it doesn't get the best deal.
44:29
So it's a very fragile situation and how you play that out in the
44:33
future is going to determine a lot of things.
44:38
So in our industry, a big part of our customers and where we
44:43
earn most money are with professional car dealers, also
44:47
professional real estate agents. And but it's the same with what
44:52
you're saying, Dmitry, But some thoughts on that.
44:55
Yes, you could have an agent that would help them, you know,
44:58
put all their inventory everywhere.
45:00
But those tools already exist today.
45:03
They have tools already and just standard API connections.
45:07
You know, there's, there's software that allows you to, I
45:10
think they called multiple listing systems in our industry.
45:13
So they have an inventory management system and they can
45:15
just select like, Oh, I want this car on, on these five
45:18
platforms, right? And and then the API connection
45:21
just handles it. Sure, you could ultimate more
45:24
pieces, right? For example, there would be
45:26
customers calling or sending a message and you could
45:28
automatically reply to them to make your selling process
45:31
easier. But in the end, they typically
45:32
want, you know, they want to stay very close to that
45:35
potential customer because that's how they can sell, that's
45:38
how they can use their skill. So it's a very interesting
45:41
domain. I've been thinking about it
45:43
whether seller agents in our context, so for example, car
45:46
dealers and real estate agents, if it makes sense.
45:48
And I'm not certain because I'm not certain what use they will
45:51
have beyond tools that already exist today.
45:55
Yeah, maybe it could scale more, you could be on more platforms,
45:58
but typically there aren't that many classifieds type platforms
46:02
in a market, right? You'll maybe have like 3, Max 4
46:04
or so relevant ones. So it's a very interesting one.
46:08
It might be different indeed for let's say a producer of a good
46:12
who just wants to sell to every market in the world, they don't
46:14
care. But when it comes to these kinds
46:15
of selling agents, I'm not certain how the how the future
46:20
will look. Definitely interesting.
46:23
Maybe it's a topic for another debate.
46:26
Well folks, I think we are coming to the end of the great
46:30
practitioners discussion. I want to thank Dimitri, Steven,
46:34
and Tiara for being here with us today to give their opinions on
46:38
what the future may look like around these two topics and
46:43
really being intellectually honest with what is real right
46:46
now versus what is a little bit more hypey.
46:50
This was a very special production of the Process and ML
46:54
OPS Community AI agents in production collaboration.
46:58
It is the 5th episode of the miniseries that we kicked off.
47:02
I was able to spend a week at the Process AI offices and embed
47:06
myself in their teams, learn about all the agent use cases
47:10
that they're working on, and then bring the different experts
47:14
into the studio with me to explain what they've learned
47:17
while building out these agents. Since it is something new, I
47:22
would love if you give me your feedback by either leaving a
47:27
comment on Spotify or YouTube or dropping in a review on Apple
47:32
Podcasts or wherever you listen to your podcast because it helps
47:36
me so much to hear the feedback from you all.
47:41
If it's too much to leave a review, just DM me on Twitter or
47:44
LinkedIn or Blue Sky. Even hearing your feedback, it
47:48
is a great way for me to convince people like the Process
47:51
AI team that it is worth their time to take the day off and
47:55
come and chat with me in the studio.
47:57
Lastly, these awesome people that I've been hanging out with
48:00
at Process are hiring, so we'll leave a link to all their job
48:04
openings in the show notes. Check it out and peace out
48:08
everybody.