Agent Workbench - Ep 1 - The AI Agent

Hi everyone, Thiago here. I'm an

outsystems architect and an agentic

architect and today we are going to

start exploring agent workbench. But

just before going there and start doing

practical things and talk about the

different parts of it. Let's start by

the definition of what an AI agent is.

And the reason I'm doing that is because

I I I see a bit of a a challenge on

everyone to know exactly what should be

the scope of an AI agent, right? Because

so far we have human agents, so far we

have automation, but now we have this

new player in the market, the AI agent

and not every use case should be given

to an agent. So let's explore what an AI

agent is and let's finish here with a

more target sense of what should be a

correct use case for an AI agent. So

let's do this

starting with a very simple question.

What is an AI agent?

And to start with this, let me first

show you here an image. So most of you

are already familiar with this, right?

So this is chatupt

and a lot of us use this uh some in a

daily basis, others when they need

something specific, right? But what I

want to ask you guys is do you think

that chachip is an AI model or an AI

agent?

And to answer this, let's start start

first by an AI model. So what is an AI

model? The AI model is the foundation of

an agent. So this means that the AI

model is a closed box with a very

complex algorithm inside to take

decisions. And this algorithm is mostly

made out of uh really complex neural

networks. uh you might heard already

things like transformers

uh and things like that. So it's a very

complex algorithm and the only thing

that he does is to receive a set of

inputs in this case tokens in a form of

text right and based on the input that

he receives he and based on the train

that he got from huge amounts of data he

calculates that the best uh or the most

suited next token to give is the word

blue. And then if we concatenate this

and give it back to him, the sky is

blue. He calculates again the next word

and gives the end. And we do this until

we consider that we arrive to the end of

the answer. So you can see that an AI

model is a static thing, is a

an isolated thing. It does not have

memory. He does not have tools in the

way that we are normally

giving them nowadays to agents, right?

He is just a very complex algorithm in

the box. And if you guys want uh there

is a really cool visualization, a 3D

visualization of uh some AI models um in

this link. I will put it also on the

description so you guys can go there and

see a bit more in detail what an AI

model is. But now let's go into the

agent. And this is a tricky thing

because

despite the fact that all of this hype

nowadays makes us think of an agent has

an AI thing, right? In reality, agents

already exist has a concept since the

70s when we started first to have

multi-agent systems uh used for example

in factories in robotics. Um so since

that moment that there was this concept

of an agent and an agent is an entity

that has the capability of reading from

the environment

of making decisions based on that and

then taking actions.

So if you have a piece of code or an

entity anywhere in the world that has

these three characteristics. So it's

something that can read from the

environment. It can make the decisions

based on some decision maker algorithm

and then take actions. This means that

you have an agent, right? So this means

that for us to have an AI agent, the

only thing that we need to do is to take

this middle part the that we have that

we call the decision maker, right? And

replaceer replace it by an AI model. And

by doing that if we complete the concept

of having different types of inputs like

user input uh giving a database uh

information or giving a web page and we

also give him the capability of in

taking actions upon the environment

right like giving responses or calling

tools to write on databases run logic

whatsoever. So this means that we are

facing an AI agent.

So

if we want to look into an AI agent, an

AI model and compare it for example with

ourselves, this means that having an AI

model is the same thing of us having a

brain. If we look it in our brains

alone, they cannot do much, right? Only

our brains or our brains alone cannot

interact with the environment. They

cannot take action upon the environment.

They can make decisions but they cannot

do anything rather than making those

decisions. Right? To do this a brain

needs to have inputs. He needs to have

eyes. He needs to have ears. He needs to

have uh nose so that he can get data

from the environment. And at the same

time he also needs to have the

capability of acting upon the

environment right he needs to have a way

to command our arms to move to touch

right but he also needs or um he can

also use the mouth for us to talk like

I'm doing right now. So if we combine

our brains with all our body including

sensors and actuators

we have an AI agent. So in this case as

humans we are also a type of agent right

because we also have a brain has a

decision maker we also have sensors to

read from the environment and we can

also act upon the environment. So we are

indeed a type of agent.

So

so far we talked about this difference

between an AI model and an AI agent,

right? So if we look again to the first

question of is chat GPT an AI model or

an AI agent. If you guys notice nowadays

when you are using CHP you can see that

he has the capability of reading

different sources of information. So you

can give him documents text but you also

can tell him to fetch web websites

information. Um right so he has

different types of inputs to read from

the environment but he also has

different outputs. He can also um call

tools. He can also generate code and run

that code. So in reality when we talk

about CH GPT we are talking about an AI

agent.

So chat GPT is in fact an AI agent

because if we realize from it it has all

the capabilities to be considered

according to a definition an AI agent.

Now

we talked about what is an AI agent,

right? But in reality, we know that he

can act upon things. He can read from

the environment. But where can we use an

AI agent? So let's look into a typical

scenario of an enterprise, right? So

imagine that we have a business and that

business uses or has a CRM application

or a CRM domain, an HR domain and an ERP

domain. And for each one of these

domains, they have a set of applications

with screens. They have background logic

and they have their own databases and a

set of APIs. Right? So you can imagine

this as being an outsystems environment

but also any other technology stack. So

this means that we can look into the

different layers that the agents can be

implemented right. So considering an

agent something that reads, thinks and

acts, right? So we can look into the

different layers and starting for

example with the layer from the UI. We

can have a chatbot for example that is

implemented in a web page or a web app

and we can talk with him using text.

This is a very common use case that we

see nowadays. But we can also have an

agent that is inside a certain process

or in on the server side in the

background

uh logic of the application and

performing all sorts of proactive

activities. He can be for example an

employee validator or an onboarding

agent here for the HR system. And then

we can also have an agent that it's not

on the database because it's not

possible to have agents on the database

but he can be specialized in reading

data fetching data from the database

with the goal of validating that data

checking the health of the data making

fixes and so on. Right? So we have here

in these three layers but then we can

also have an agent coming from the

outsides right because the fact that we

expose APIs to the exterior this means

that we might want also to give the

capability of external agents to act

upon our systems right so we can also

have an external agent contacting our

systems and at the same time we can also

have an agent and fetching data from our

systems to produce other effects outside

the scope of these three applications.

So when you see we can indeed deploy

agents in a very huge stack of

possibilities. We are not just talking

about having a cute chatbot uh on the

the web pages. We are talking of full

stacks of agents that can really

automatize, improve and make our life

better.

So

we talked about what an AI agent is. We

talked about the places where he can be

put. Now let's see what type of tasks

should we look into it. And the reason

to do this is because not all agents

should be treated in the same way. Not

all AI agents are a generic universal AI

agent. It always depends on the use

cases. And if we look into the three

aentic layer coverage, you can see that

it is very common in a business that

converts or wants to convert into an

agentic uh architecture. It is very

common to have these layers of agents

where we have personal assistants that

are really close to the user and they

are uh more targeted to uh talk with the

the with the user and to build a

relationship with the user. So they need

to be uh more conversational models. But

we then have the business agents and

these business agents they are indeed

experts in a certain domain and they

should work under the hood in the

background uh improving things

automatizing and so on. And these two

agents they are completely different in

the way we build them. And if you want

to know more about it, you can check my

previous videos about the the the three

agentic layer coverage and the way we

can build these agents.

But now

going again to the tasks

there was there was a study performed by

Deote in 24 that they uh evaluate

hundreds of thousands of tasks in

governance and their goal was to try to

classify the tasks based on what task

should be mostly given to humans

AI or automation

And so they analyzed the tasks and they

created three dimensions on those tasks

and they started with the first

dimension of creativity difficulty. So

the meaning the capability of towards a

certain task be creative about it and

come up with new solutions new ways of

completing that task. Then we have the

context variability. And this context

variability is the capability for us or

the agent to adapt

the to the task to the environment and

perform accept changes or variations

that might be required to complete that

task. And finally, precision. So the

effectiveness

or the accuracy that you can complete a

certain task using the less amount of

resources.

So in this triangle

what we notice is that before this new

era of AI there was a really clear way

of separating human work from machine

work. So if we were talking more about

precision, it was all about automation,

right? Making things work almost 100% of

the time. But if we were talking about

tasks that require creativity or uh

variability, adaptation, we were talking

about human domain. But then with the

evolution of AI, we start having more

and more modules and more and more uh

complex algorithms that could approach

and could give some creativity

capabilities and s some adaptation

capabilities. And then

large language models appeared. And this

completely shift the way we looked

before into human work and machine work.

Because with large language models and

the different types of models that

started to appear underneath it, we

start noticing that they were really

useful for tasks more on the human work

side. And this brought a new era to gen

generative AI. So this means that in the

initial triangle

defined by these three uh dimensions of

a task, we can give automation or we can

give tasks that require really good

precision mostly to automation. So we

are talking about um things that cannot

fail. um if they are repeated 1,000

times they cannot fail. So we are

talking about robots. We are talking

about deterministic codes, deterministic

algorithms, right? But humans they are

they still domain the area of context

variability and creativity. So for jobs

that rec require really social

interaction

um or really complex physical tasks,

those still are still done by us human.

But then we have here this middle of the

triangle and here is where tasks that

require somehow of a precision but they

also require some good adaptation and

creativity they lie into gen AI or

generative AI use cases. So we are

talking about summarize, we are talking

about uh compare specifications

uh comprehensive

tasks, explain uh subjects uh and for

example generate codes and functions. We

saw that AI is handling that really

well. So this is really important when

we are thinking about the use cases we

want to solve by building agents in the

agent workbench because we should always

ask ourselves if the task that we want

to solve if it is a task that requires

adaptation and creative creativity or if

it is a task that is just precise is

deterministic and there is a mathemat

mathematical way of doing it without the

need of adaptation

and this could change completely the way

we look into a use case and know if it

is a use case for an agent or not. So

let's now look into some examples

to see if you guys can uh answer

correctly if it is a good use case for

AI or not.

So let's check the first case. I want to

build a ticket support agent that reads

help desk emails and based on those

emails he can create the tickets on the

ticketing system so that he can fill the

title, the date, the priority, who asked

for uh for help and the description.

So we have an agent that reads an email,

digests that information, and then

using somehow a tool or logic, he

creates a ticket with a defined

structure.

If you answered this has a good use case

for an AI agent, you are right because

the most important characteristic here

is that we are asking the agent to read

something that is unstructured

that is the email. Right? There can be a

lot of different ways to write an email

about a ticket asking help. So indeed

the agent needs to adapt the the

situation to the text that he is

reading. Right? So this is a really good

use case for an AI agent. But also a

very important thing is that we are

asking him to read something

unstructured and then convert to

something structure. Convert to a sort a

sort of a structure. If we look into our

systems that has a title attribute, a

date, a priority uh request and a

description. So this is also a very good

use case for uh AI, the conversion of an

unstructured data to structured data.

Now let's look into the second case. I

want to build a data validator agent

that receives a JSON

and checks if the fields on that JSON if

they have the correct data format and if

they have for example if it is a text if

they have a length below 500 characters

and if the validation of that data went

well he stores that result the JSON in a

database.

Is this a a an AI use case or not?

If you answered that this is not an AI

use case again you are right. And the

biggest reason we are considering this

is because all of the actions that we

are asking the agent or not to do, they

are deterministic. So let's see for you

to check if a certain attribute

has the correct data format. You have

lines of code that can do that. Check if

it is a text. check if it is a number.

Right? Also to check if a certain text

field has a correct length,

it's also a deterministic line of code.

You use in case of out systems, you do a

length of the fields and you check if it

is below a certain value.

And then finally, if everything is okay,

you are saving a record in the database.

that is also deterministic because you

are using a crude action to do that. So

you see all the actions that we are

asking here for the agent to do they are

deterministic. They don't need any

adaptation. There are lines of code that

can do that directly. And because there

are lines of code that can do that

directly that is more effective because

that is more precise. It's it's faster.

It requires less processing power. So it

is better to simply use automation.

Let's now check the third use case.

So I have or I want to build a medical

advisor agent. So I want an agent that

picks the full historical data of a

certain patient and based on that data

and based on all medical knowledge in

terms of the of all books all sorts of

information that exists he builds a

report and he sets a group of actions uh

treatments and appointments for that

patient. Is this a use case for an agent

or not?

So, this one is more tricky because to

simply start this with a yes or no,

we have a yes, but the reason for that

is because this is a very huge use case.

Just imagine for example this scenario.

In this first case, if you wanted to

train a human to be able to do this,

most likely in one week you would be a

you would be able to have someone have

already a human doing a good job in here

in picking an email and then filling a

ticket. Right? So we are talking about

of one week of on boarding right and the

reason for that is because the tasks are

quite simple or the amount of

information is not big. In this third

case,

if you wanted to have a human to be able

to do this,

they exist nowadays. They are called

doctors and they need to study years and

years and they need to continue be and

be continuing learning new uh skills and

new information because this is a very

huge and complex task to do, right? for

you to be able to check all the history

of a patient. So we are talking about

reports, we are talking about

information from exams and um even for

example if there is IoT devices uh like

a heart device. So we are talking about

a lot of data but also about a lot of

medical knowledge right if we see all

the books all the

gigabytes of information we might be

talking in terms of medicine it is a lot

to process it's a lot of text so this

means that yes you can build an agent to

do this but it will be very difficult

for you to have an agent

that is really effective in this

situation. It will be really likely if

you just give him the information and

ask him to do something and uh receive

the results from it. It's very likely

that he will not be effective. He will

not give you good results in terms of

quality. So how do you fix this? You

break into smaller use cases. So you

could have for example an agent that is

only job is to index the huge amounts of

medical data that exists from the

outside and to use for example a rag

system to retrieve contextualized

information and then to summarize it

that information to a second agent and

then you could have a second agent that

receives the data from a patient

receives gives the summarized

information from the other agent and

then based on that probably he can

already do a better job. But even so it

might be required for you to even break

this into a bigger team of agents. So

yes, it's possible but it's not easy.

Fourth case, build an a an ebam

validator agent. So, I want to have an

agent that picks a set of transactions

from a certain business and validates if

the eban is a valid number and if so, he

updates the status of the transaction to

verify it.

Is this an AI agent use case or not?

So

if you took the example from the second

case and if you know in fact how an EBAN

is verified, you might notice that no,

this is not at all a use case for an

agent. And the reason for that is

because the EBAN validation is a

deterministic piece of logic. So it

follows certain rules. He sums numbers.

It checks characters in certain

positions, but all of those operations

are deterministic. So this means that

most likely you already should have a

library, an orange ball somewhere that

could do that in the deterministic way.

Finally,

the last case, a similar one, but let's

check here. I want a financial

transaction agent that again picks a set

of a transactions from business and

verifies if the data is correct of those

transa transactions. So if it has an if

the is valid uh if it has a date um if

the date is in the past if it has a

description and so on.

And he also verifies in the description

if there is enough information to track

uh track that transaction to a certain

operation to a certain order. So it

checks if the it has the necessary

details about the product and the store

and other details that might be

necessary and if so he puts the

transaction into the status of verified.

So in here what should be the correct

answer? Is this a a good use case for an

agent or not? This might be similar to

the use case for but there is a small

difference is that we are asking the

agent to verify the description and

check if there are some conditions in

that description. And that description

is in an unstructured information,

right? Because we might have a store

that puts first the product name, then

the the name of the store. We might have

another store that puts first the name

of the store uh including the address

and then the product name, full name.

You might have another store that just

checks uh if or just puts a product code

and the name of the store. So there can

be a very very different uh very many

different and structured ways of filling

the description. So in this case we do

need adaptability right. So just because

of this small condition this means that

we cannot do all of this using just

lines of code and deterministic

u functions. We need indeed to have an

agent and yes there will be a mix of

validations that can be done by code and

should be done by code. So the agent

should have the capability of giving

that information to an orange ball to a

certain logic to verify what it is

deterministic. So to check the check if

the fields are in the correct format and

so on. But then he also needs to do on

his own not giving to a certain piece of

code he needs on his own to validate if

the description is wellmade or not.

So this is indeed a use case for an AI

agent.

So after these use cases, let's recap

here what we did. So today we talked

about the difference between an AI model

and an AI agent. After that we picked

this concept of an AI agent and we

aligned the different layers that we

could introduce AI agents in web

applications, right? So we are we could

be talking about the UI layer logic or

more close to data or even to allow

consuming or exposing APIs for agents.

But then we finished this by also

analyzing the different types of tasks

that we can have and what are the better

tasks that are suited for AI use cases.

I hope that this video could help you um

depreciating the different scopes that

we can have and this is really important

because

due to that if we deliver

uh better AI use cases, more suited AI

use cases, we ensure to our clients and

to our projects, we ensure that

outsystems stays a good quality AI

powered low code platform that can

deliver not only applications but

agentic functionalities to the business

uh and the issues that we want to solve,

right? But at the same time, we are

starting to have a lot of agents in the

forge, right? Everyone is doing uh their

part to start working with the community

and start working together developing

agents. So it's also really important

that we start from the beginning knowing

what should be a real concept of an AI

agent so that we can have indeed better

quality of agents in the forge. I hope

that you like this video. Let me know in

the comments if you have questions about

this and have a great day.