YT YouTube2Text

YouTube2Text → Transcripts → Discover Agentic AI | Workshop 1 | Create AI Agents Championship

Discover Agentic AI | Workshop 1 | Create AI Agents Championship — Transcript

by Create AI Agents · 15,491 words · 2,192 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:09My name is Sarah Saduk and I'm part of
0:11the create AI agents championship. Some
0:15of you have been here with us last year
0:18for the create apps championship and
0:20this year the create AI agents
0:23championship is an extension a new
0:25version of the championship itself. So
0:28welcome back for those who were there
0:30last year and welcome to those who are
0:32new. This year we are focusing on
0:35teaching you what is an AI agents and
0:38how to build yours.
0:41>> Next.
0:43>> So before we start if you want to speak
0:46just uh raise your hand um ask your
0:50question on the chat let us know we will
0:54be there to answer your question. The
0:55chat is open and there is a Q&A at the
0:57end. Respect the space of others as well
1:01on the chat itself. So um and feel free
1:04to interact with each other.
1:15>> So in case of uh you are joining from
1:17your laptop all the links and shared
1:19resources can be accessed from your
1:21computer. Uh you can keep your phone
1:23with you as well because we're going to
1:25be asking you to uh scan some QR code
1:28during the session and please uh rename
1:31yourself using this format your name and
1:34the name of your business or of your
1:37app.
1:49Okay, we have a small uh thing to ask
1:52you. If you haven't joined the WhatsApp
1:54community, scan the QR code and uh join
1:58us on the WhatsApp community. We're
2:00going to be closing the WhatsApp chat um
2:03the uh normal chat because we are out of
2:06capacity. We have too many people. So we
2:09are going to only keep the WhatsApp
2:11community. So feel free to scan it.
2:17Let's uh go back to the competition
2:18itself. The create AI agents
2:21championship has been announced by his
2:24highness shean bin Muhammad bin Rashid
2:26al-Maktu who is the crown prince of uh
2:29of Dubai and it's led by the Dubai
2:31Chamber of Digital Economy.
2:34The create AI agents championship bring
2:36together beginners but also experts who
2:40have an idea of an AI agents
2:43and want to compete for prices. We are
2:46going to be providing you mentorship
2:48during the session. I will go through it
2:50step by step. Um you can enter the
2:53competition either as an individual or
2:56as a team. For that I will ask you to uh
2:59to register all your team member in the
3:02same way you registered yourself and
3:04from November onward we are going to
3:06open some uh session of in the digital
3:10academy space and you will be able to
3:13learn on your own how to
3:16develop an AI agent.
3:24the the competition has different
3:27prices. Four prizes uh four categories
3:30that you may have seen on the website
3:32and another category will go through it.
3:35Most importantly for now we are asking
3:37you to attend the three workshops that
3:40will be running until November and then
3:43we will have another series of
3:44workshops. So at the moment they are
3:46pre-ompetition workshops from now until
3:48the 4th of November. Next,
3:54in terms of categories, we have four
3:56different categories. The best social AI
3:59agent, the best personal AI agent, and
4:02um the best business AI agent, and the
4:05best youth AI agent. The last category
4:09is for people who are under 21 years
4:11old.
4:13Each category has a price of $150,000.
4:18And last but not least, we have one uh
4:22category that is the Dubai AI agents
4:25category of the year. And there you you
4:28will earn up to a million dirham which
4:30is equal to $272,000.
4:33Now these are price funds that you will
4:36have to invest in your business. Next.
4:44Next.
4:47>> Right. Let's go into the timeline
4:48quickly and then I'll give the floor to
4:50Moral for you to learn how to uh
4:53discover how the AI agent world. So the
4:57first phase of the competition is about
5:00um registration of different
5:03participants. During this registration
5:05phase, we are providing you with three
5:08introductory live workshop with industry
5:10expert on house under the uh live
5:14workshop. You can register to those
5:17three workshops. Then we will be opening
5:20the digital academy in November. And
5:23last but not least in terms of
5:25competition training, the eight uh
5:28workshop and master classes will be
5:31happening in November and December. In
5:34December, we are uh going to enter the
5:38submission phase and in January, we'll
5:40be requesting you to submit your AI
5:42agent, your pitch deck, and your demo
5:44video. If you are qualified to be a
5:47semi- finalist, you'll be entering
5:49automatically an accelerator program
5:51that will help you develop further your
5:53AI agent. By the end of the accelerator
5:56program, what we are uh requesting in
6:00the in in the submission phase is for
6:02you to have a live AI agent. This is the
6:05only way if you're a semi- finalist to
6:08uh to qualify as a finalist itself
6:11because during the final ceremony that
6:13will be happening in Dubai, we are
6:15requesting you to showcase a live agent,
6:19of course, a pitch deck and a demo
6:21prototype of your agent. We want to
6:23understand what it does, what is it
6:24solved, but also how it works and how it
6:26look. So the pitch deck will include few
6:29elements that is are related to your
6:32idea itself, the competition, the the
6:36market size, the customer, everything
6:39that is related to the business. And
6:41this deck will be provided to you uh in
6:44November when we start the competition
6:46training. The semi-final happen in March
6:49and the finale will happen in May
6:52itself. For those who do not make it to
6:54the um semi-final
6:58uh they are they have the possibility to
7:01be part of the high potential team which
7:04are um a number of team that we identify
7:08as teams that have a high potential of
7:10development uh in the region but do not
7:13qualify as semi-finalists. We'll go
7:15through it step by step. But this is to
7:17give you a glimpse of what the
7:19competition is all about. Without uh
7:23further ado and before I introduce you
7:25to moral, I also want to let you know
7:27that tomorrow like every every Thursday
7:30we have a checking call with you. So it
7:33will be an open meeting where we opening
7:35the floor to to the participant to ask
7:38us question about the competition. So
7:40join us. It's tomorrow. It's from
7:443 to 4
7:46>> and the link is on the house space on
7:49the community chat but also in the
7:51workshops part part. Without further
7:53ado, now time to learn
7:57uh time to learn about the
8:01uh the houseace. So the houseace as you
8:05have seen is a platform where you will
8:08be connected with the community and have
8:09access to learning. is very
8:11straightforward and every time we'll be
8:13opening a new page. So, uh feel free to
8:16interact with others and to keep an eye
8:19on this platform.
8:30All right. Now, I'm gonna introduce you
8:32to Moad Al Mazyani who is part of the
8:35create AI agent team. He is a tech and
8:38digital consultant. He's been working
8:41with private sector, public sector and
8:44he's a figure in the agentic AI space.
8:48Welcome Mural. We are very happy to have
8:50you here with us.
8:52>> Thank you very much Sah. Thank you very
8:54much everyone attending. Very happy to
8:56be here and very happy to uh to
8:59facilitate this session today and to
9:02share a lot of knowledge.
9:05So uh without I will be sharing my
9:09screen.
9:12Okay.
9:18Great.
9:22Okay. I will be introducing myself. My
9:24name is Morad Miziani. I'm um consultant
9:27um since 18 years in the digital uh
9:32area. I've been working with uh with the
9:35projects uh in private sector in public
9:38sector and uh we deployed u multiple
9:42agents and also uh multi- aents uh
9:46applications lately and uh today I will
9:49be sharing with you uh how we we do that
9:54uh and a lot of uh ways of building your
9:57own agents. So uh we will have live
10:00demos. I will be sharing how we can
10:03build uh different kind of uh agents.
10:06You already have your ideas but we will
10:09share best practice in the sector.
10:13So um before starting I want to um ask
10:18you some questions about uh what will
10:21you delegate to an agent? What will you
10:24keep human and uh what do you think? It
10:27depends. So I'm curious about having
10:30your answers if you can scan this code
10:33and we will play this game on uh this
10:36platform to know your response.
11:01Sh.
11:04Okay.
11:15Okay. So, we can see your answers.
11:20I can see that uh already you are
11:23sharing a lot of stuff with your AI
11:25agents.
11:26So for first question booking a table
11:30some want to keep it human and uh many
11:33we want to delegate this because uh it
11:36can be uh it can take a lot of time. It
11:39can u take time to compare to to to do
11:42some bookings to call to to to know if
11:45there is some availabilities.
11:47So basically this will be something we
11:49can uh eventually uh delegate to an
11:53agent. Next question.
12:01Okay. Now let's speak about friendships.
12:05Do we want even our friendship to be
12:09managed by AI agent? Wow. I can see that
12:13multiple persons uh responded that they
12:16will delegate
12:23but uh obviously we keep it human. Okay.
12:28But I can even see 35 persons they want
12:31to delegate to AI agents.
12:34Keep it human. Of course, friendship is
12:37uh what uh make us humans and what
12:40connect us and uh it's one of the things
12:43that uh we hopefully want to keep human.
12:47Okay.
12:49[cough]
12:50[clears throat]
12:52File your tax returns. interesting
12:54question and um
12:58I'm curious to know
13:03if you will delegate and verify. Thank
13:06you very much. Delegate and verify. So
13:09we consider our agents like interns and
13:13it can be very important very
13:15interesting to delegate some tasks but
13:18uh here we are talking about tax return
13:22and uh a mistake can cost us a lot of
13:25money. So delegates and verify.
13:30Interesting answer.
13:33Delegate by AI, check by human.
13:39Thank you. Next question.
13:43Okay. Now we have an angry customer
13:46complain.
13:51Keep it human. Interesting.
13:59It can be both because like an AI agent
14:02can be faster to respond. But of course
14:05we are speaking about the human angry
14:07customer and as business owners or app
14:12owner or agent owners we need to stay
14:14close to our customers.
14:21Good.
14:24Sure.
14:25Keep it human.
14:28Okay. Choose the outfit for tomorrow.
14:30This is a personal one.
14:39I can see that many people want to
14:41delegate that to AI agents.
14:48Interesting.
14:51But multiple want to keep it human.
14:58Thank you very much. So I know you
15:00better. What do you want to share? What
15:03do you want to delegate? What do you
15:04want to keep? Very interesting to know.
15:08Um because like agentic
15:11agents are part of our teams, agents are
15:14part of what we are doing. So you
15:18received yesterday a form that you
15:20filled and uh we received uh more than
15:23100 answers and you shared some ideas
15:27about uh what do you know about AI
15:31agents? How what do you know about
15:33connectors? What's the difference
15:35between an assistant and an agent? Uh
15:39how can we share access with agents? And
15:42of course how to approve what an agent
15:45is doing. So from these forms
15:50um many of you uh have uh like uh for
15:55this question if uh I move if this app
15:59moves a meeting from Friday to uh
16:02another day is it an agent or is it an
16:05assistant or is it an automation? So 64%
16:09say it's an agent. We will see why uh
16:12what's the difference between the chat
16:14bots that we had uh from for years and
16:18the difference with agents that we have
16:20now and they can do actions inside of
16:24tools. Um we also asked you about the
16:27the the categories you are going to uh
16:30push your idea to and um 47% are doing
16:36ideas for business operations finance
16:39sales. Uh we have uh social 18% personal
16:4415% and 15% from you are not sure yet
16:49and uh we have only 5% of youth. Uh we
16:54really want more youth. Speak about this
16:57program to your friends to your family.
17:00This is really important to have more
17:02youth
17:03for ideas. This is the the the ideas
17:06that uh was collected. So we see another
17:09time business operation sales human a
17:13lot of human this uh this this key word
17:16has been repeated lot of time WhatsApp
17:18customer uh automating tasks autonomous
17:22agents health payments so we see uh that
17:26there is multiple
17:28multiple use cases uh with the business
17:33so um some ideas about what an agents do
17:38and uh what what you want to do. So we
17:42have finance, we have personal life and
17:44productivity, sales, learning and
17:47education. We has also uh ideas from uh
17:50security, IT and software. This was your
17:53your your ideas and we asked you in the
17:57form to share with us your uh your
18:01prompts or your instructions for the
18:04agent. So this was a first uh knowledge
18:09for us uh to know how do you give
18:11instructions to agents and we collected
18:14some ideas some uh some examples from
18:17what you shared with us from example
18:20this example is uh I'm giving to my
18:23agents please make a marketing campaign
18:26for this company with this budget target
18:28is Y and Z and this is what I want to uh
18:32know. So uh basically you can see that
18:35this prompt is very basic. Uh it has a
18:39goal and uh it's uh it will pretty work
18:42with uh with the chbt with code with
18:45those agentic apps but to have an agent
18:50that can be u [clears throat] done for a
18:53business that can have uh some some
18:56goals and can can give some great
18:58outputs. It's missing context. It's
19:01missing tools and permissions. It's
19:04missing the output formats. It's missing
19:07the human approval. We spoke about it
19:09and it's missing the checks. So this is
19:12the best practices for uh an agent to to
19:15to have the the the the best output
19:18possible. So today this prompting this
19:21instructions are like source code. Uh we
19:24are giving a lot of context to this
19:27agent. uh not much context but uh lot of
19:32good context for this agent to give the
19:34right answer and why we are calling it
19:37an agent uh because uh the LLMs are the
19:42brain uh of the AI but agents are the
19:45hands of the AI and they can perform
19:48actions inside tools. So they can read
19:52for example our agenda. They can go to
19:54the CRM and bring some informations.
19:57They can send emails. They can do some
20:01actions. And uh when we give permissions
20:04to an agent, he knows that he has access
20:08to these tools and he can use these
20:11tools to give the best output possible.
20:13It can even go for some financials. For
20:16example, we have an agent on WhatsApp
20:19and we want uh we want this agent to
20:22send for example quotations and we can
20:25give our agent access to our our uh our
20:29uh software uh and it can create this uh
20:34quotation and send it. So without having
20:36this access it cannot perform actions
20:39and um we are not going to call it an
20:42agent because an agent without tools and
20:45permissions is simply an assistant who
20:48can answer to to questions and it's a
20:51chatbot. Defining the output format is
20:54also interesting because
20:57uh we we want our [snorts] agent to be
21:01to to have the best output possible and
21:03also when we are building an agent we
21:06want to optimize the tokens. So instead
21:09of having output one and two and the
21:12user will say no this is not the right
21:14output. We need to define the right
21:16output so that uh we we can optimize the
21:20token consumptions and we can also
21:23optimize the the the uh the exchange
21:26with the with the the final user human
21:29approval. This is what we are calling
21:31human in the loop. Uh today when we have
21:34super agents uh that can perform
21:37multiple tasks for example in accounting
21:40uh we can say okay you can perform this
21:43task you can automate this task but
21:46whenever the the scoring of this uh of
21:50of what you are doing is below for
21:53example 9.9
21:55uh I want to validate that this is what
21:57we call human in the loop and it's very
22:00important in your agents when You are
22:03doing as the example we saw, you are
22:05doing some tax agent, you are doing
22:07something that have uh legal uh
22:10approach, you need to have human in the
22:13loop and checks is also very important.
22:17>> Wait,
22:19can you check the the chat? I'm I just
22:22want to send you a message quickly.
22:24Sorry to interrupt you.
22:26>> Yes. Yes.
22:30So the chat here.
22:38Yeah. Uh we are still in the slide one.
22:41I'm still speaking about the slide one.
22:43So uh I'm I'm speaking about the best
22:45practice.
22:46>> Perfect.
22:47>> Yeah. Speaking about the best practice
22:49we have here. So [clears throat] we have
22:52this prompt, this simple prompt, but an
22:54improved prompt will be something like
22:57that. And even more than that we will
22:59see uh some improved prompts. So we were
23:02speaking also about checks. So checks is
23:05the checklist and I tell to my agents
23:08before giving an output uh this is the
23:11checklist you need to to verify. So this
23:14is also a best practice uh to have the
23:16best output possible in our agent. So
23:20here is another example from what we
23:22received yesterday. uh the instruction
23:24for your agents plan 5 days vacation in
23:28Dubai and uh this is uh make sure you
23:31pick the closest hotel uh and book it.
23:34So this is also here we have the goal
23:37checked but we don't have the other best
23:40practices checked. We don't have the
23:42context we don't have any tool or
23:45permissions we don't have the output
23:47form. Uh of course it's not necessary to
23:50always have all the best practices. For
23:53example, here we will not necessarily
23:55need the human approval, but we
23:58obviously will need some tool access uh
24:01to to make it more accurate for example
24:03and also an output format to have the
24:06best uh format each time and not iterate
24:09in the formatting. Another prompt uh
24:12with the goal. So what we saw from your
24:14exercise are a lot uh from participants
24:18that fill the form respond to goal but
24:22uh they are missing the other uh best
24:24practices. Uh sometimes we have the
24:27context and uh the best the context we
24:30have uh having having more context
24:33doesn't mean having the best context. So
24:36uh for example if I have a legal
24:39assistant or agent uh it's not by giving
24:42all the legal documents that it will be
24:45a better agent but by by creating the
24:49best uh legal documents for a niche uh
24:53it will be better. So today we have uh
24:56code we have chpt that can perform very
24:59well in legal for example but when we
25:02are going to think about an idea we need
25:05to think about a niche and serve it with
25:07the right context and uh if we have some
25:11data that is not available in the
25:13internet it's even bit better
25:16another uh prompt improved uh so this
25:20time we have goal we have context we
25:22have tools and permissions
25:24So we can uh give some tools and
25:26permissions. We have human approval.
25:29This is a good one. This is a good
25:30prompt. But of course uh as we said the
25:33prompt engineering and uh the context
25:36engineering right now is uh we will see
25:39some uh real examples is like source
25:42code. You you need to go into details
25:44into each situations and how it works.
25:49Good. So uh I want to ask you some
25:52questions because not everyone uh shared
25:55their uh the form yesterday only 100
25:58people. So I want to know you better and
26:01uh if you can scan this mentimeter link
26:05and fill the form please.
26:49Okay, good. So, um
26:52um today AI um is part of our daily
26:56life. It's like Excel, it's like Word,
26:58it's like using the internet. So of
27:01course uh a lot of people uh are using
27:05uh those tools in everyday uh decisions
27:09but uh we also have a lot of people
27:12building their AI agents and this is
27:14really interesting. If we can move to
27:16second question
27:20about the categories
27:33business.
27:48Okay. So, it doesn't change from
27:50yesterday. Many of you are um building
27:54business uh agents.
27:56So I can um guess that uh it's about
28:01operations
28:02uh and uh
28:05many many operations automations and
28:08agents. Okay, let's go to next one.
28:13Do you have prototypes? Do you have
28:16ideas? Are your agents live?
28:32Okay. So prototypes that's uh that's
28:34great. So we have multiple prototypes
28:37already ready
28:39and uh even some live agents and uh a
28:43lot from you have ideas. Um so I hope
28:47this session will uh will help you clear
28:50your idea and move to prototype. Uh so
28:54today it's easier to move to prototype
28:56with no code tools with coding tools and
29:00uh we will see how we can uh go ahead
29:03and go from idea to prototype and then
29:06from prototype to life. Thank you.
29:12Okay.
29:14Um
29:16so today we will see four demos and uh
29:20it will be we have multiple ways of
29:22building the agents and uh we choose the
29:26most visual one uh it's no code agents.
29:30Why no code agents? Because you can
29:33understand better how it works visually.
29:36Uh we suggest uh from for for
29:40prototyping each time like to have a
29:43workflow uh to to to know what your
29:46agent and sub agents if you have some
29:48complex app how it will work or if you
29:51have an agent how it will give take the
29:54context how it will perform some
29:57actions. So we will visually see all
30:00these agents working. We will start with
30:03a simple example of a personal AI
30:06assistant uh that can perform some
30:09actions and uh and uh respond to to some
30:12questions. And we will take the same one
30:15and make it a little bit more
30:18sophisticated
30:20uh and see how we can move for from a
30:23simple agent to more sophisticated
30:25agent. We will also uh have some
30:28business ideas uh about uh some AI
30:32business ideas that can help uh in the
30:35business area. So uh one to to detect
30:39tender AFP's opportunities and another
30:42one to watch a market uh to know what
30:46the competitors are doing to have some
30:48intelligence about new ideas, new
30:51products. So it can be used by your own
30:54uh by yourself for for your own project
30:57and it can also give you ideas uh to how
31:00to enhance your business uh AI agents.
31:04So the first one I want to share with
31:06you is my personal AI assistant. Um and
31:11uh for this I will be using u a tool
31:14called NHN. I think that many of you
31:18have heard about NHN or are prototyping
31:21with NHN.
31:22Um so we choose NHN for two reasons. For
31:26the first reason uh we can move from uh
31:31NHN we can move from it because like NHN
31:34is open source so we can install it in
31:37our own computers and uh it's easy to
31:40prototype with NHN because it's all
31:43about uh nodes. So, NHN means node
31:47infinite node. We can connect multiple
31:49nodes. And uh one way of building agents
31:53with no code and low code platforms uh
31:57is NHN and they have also a lot of uh
32:01nodes that can be very helpful to
32:03prototype our agents. When we are going
32:06to move, if we have a scalable app or we
32:09want to scale our app, of course, we
32:11suggest to go on full code uh
32:14application and uh you can also move
32:17easily from NHN to full code because you
32:20have everything on H nhn you are doing
32:24uh can be exported in JavaScript and can
32:27be a good start uh to have your your uh
32:31specs to start your agent. [snorts] So
32:34how this agents work? So before starting
32:38an agent, we need to know uh what is the
32:41the
32:43what will be the first task. So here the
32:48how my agent is triggered. Uh here I
32:53have a trigger. I say okay this agent
32:56wakes up every day I at 8:00 a.m. It has
33:00a goal. All agents have a goal. So we
33:03have a trigger, we have a goal. Uh the
33:05goal is the instruction. So um I say
33:10okay you are more at personal
33:11assistance. You are connected to tools.
33:14As I said we are not speaking about
33:16agents if we don't have tools. The
33:18agents are have hands to to perform with
33:23their brain to perform into tools. And
33:27of course we can have uh access to these
33:31tools. We can give access to these
33:33tools. For example, I saw many uh
33:36companies that create an email for the
33:39agent that create a team's uh account
33:43for the agent that create a WhatsApp
33:45number for the agent uh that have agents
33:49inside their Slack as well. So the
33:51agents are part of the team. they have
33:54their own account, their own access. Why
33:57it's important? Because I don't want to
33:59share my own accounts and my own uh
34:02permissions with the agents. Maybe I
34:04have uh an admin permission and I want
34:07my agent to have only some some really
34:10restricted um permissions. So for
34:13example, I want to give to my agent
34:15here. I'm doing a simple example. I'm
34:18giving access to three tools. I'm giving
34:22access to emails. I'm giving access to
34:25agenda and I'm giving access to uh an
34:28excel sheet. So here uh I will see how
34:34it work here.
34:37Okay.
34:39So I have a to-do list and uh we will
34:42suppose that this is the to-do list we
34:44are working uh we have in our project.
34:48uh and I give access to to this agent to
34:52my to-do list.
34:54I give access to an email. For this one,
34:57I have a sample email and I give access
35:01to my agenda. So, an agent is a brain, a
35:06memory, and
35:09access to tools. And of course, it has a
35:12goal and instructions.
35:15Um here I used for brain I used GPT. I'm
35:20not like uh one of the things you will
35:25face is what we call phenops and
35:28optimizing your tokens usage. Uh you
35:33will not always use the best model in
35:35the market. some uh models some
35:38opensource models are really cheaper
35:42than uh the the the frontier models. So
35:46for example here I didn't use the best
35:48model even if I choose GPT I used only
35:51GPT4 5.4 for mini because my tasks is
35:56not complex. I just want my agent to
36:00daily collect my data to analyze it and
36:05to send me a report
36:08and I use the fallback model. This is
36:10really important because you can't rely
36:12on one model even if it's uh frontier
36:15models. Uh frontier models are mostly
36:19always up but sometimes they have uh
36:22some issues as well. Uh we suggest that
36:26you always have a fallback model and
36:29also that your prompting is uh is
36:34working with the least intelligent model
36:37so that when you change the model it can
36:40still be working. And of course when you
36:42are going to use a frontier model uh the
36:45best models of course you don't need uh
36:48to go in details in prompting it will
36:50understand but given a good prompting
36:53with uh the best practice that we shared
36:55with you uh will help you uh even change
36:59the brain and also always have the same
37:02output uh and be model agnostic.
37:07The memory is very important uh because
37:10an agent have a context. Okay. So the
37:13context here is bringing it it's
37:16bringing it from your tools. But it also
37:19need to have a memory so that uh it can
37:23remember what it's doing. What was for
37:25example in this example uh what was the
37:29report that it sent to me yesterday. If
37:31I had something on my to-do that I
37:34didn't advance uh into or some project
37:37uh that has been uh for for two weeks
37:40for example uh sending me in the
37:42reports. So it can uh remember it inside
37:45the memory. So the memory is different
37:47from the context. So the context here
37:50it's the the the content here and the
37:53memory is what we put inside the the the
37:58what we put inside the database or a
38:00file. It can be a simple file or it can
38:03be uh sometimes even a database. So very
38:06simple trigger. We have uh we we have
38:11the prompt the system prompt. We have
38:14the models the memory and the the tools
38:18and we can execute our agent.
38:22Okay. So each day at 8:00 a.m. this is
38:24what will happen. So my tool my agent
38:28understood what tool it will use. Uh and
38:33um for some more complex agents we are
38:37going to we will see how how how many
38:39tools we will give to to the agent and
38:42it will decide what tool uh it will use
38:45for different uh use cases.
38:50So let's check the email that we
38:53received.
39:05Okay. So, I received this email
39:09and um this is pretty good and uh it's
39:13from my context. I don't know if we have
39:16uh some questions.
39:23Yes,
39:23>> actually we do have question on the
39:25chat. A lot of them.
39:27>> Okay,
39:28>> so some of them are related to the
39:30competition itself.
39:33Let me check those.
39:36Um
39:43no
39:44there is one uh are there restriction on
39:48the text stack model or platform can you
39:50use a third party API and commercial AI
39:52model I guess this is for the platform
39:55you there is no restriction on the tech
39:57stack we'll answer all the question
39:59related to the competition tomorrow but
40:01if you have any question for Morad on
40:03the presentation please uh send them on
40:06the Q&A and we'll address them.
40:11Yeah.
40:15>> Okay.
40:18So, I have a question about the memory.
40:20I think I answered this. Uh so, as I
40:24said for for an agent, let's go um here.
40:33Okay.
40:35So an agent uh have some goals and
40:40instructions.
40:41Uh for example uh I have uh for for this
40:46simple example it's a daily report. It
40:50can be a project report. It can be
40:53there's a lot of needs uh in businesses
40:57to have daily reports.
40:59Um so let's suppose it's the same goal
41:02in instruction. So the AI agent have to
41:06reach its goal each day. So it has a
41:09model. We say that the model is the
41:10brain and of course we have a lot of
41:14models uh from frontier models that
41:17means GPG cloud or even Gemini to open
41:21source models for example DeepSc for
41:24example Quen uh and other models and we
41:28are not going to use all the time the
41:30best model. We can use deepseek. We can
41:32sometimes use for example if we uh if we
41:36we have let's suppose we have the
41:39content it's uh some some uh paper
41:42writing uh some handwriting we can even
41:45use uh a specific model that does only
41:48that we have for example deepseek OCR uh
41:52who is specialized in this OCR and
41:54analyzing the handwriting and it's
41:57cheaper much cheaper than GPT or others
42:01so the model we always uh we choose the
42:04brain. What is a memory? So the memory
42:08is what we decide to put inside uh a
42:12file or a database that can help in in
42:16uh in the future. For example, when we
42:19do reporting, we can say always add we
42:23we can say in the instructions always
42:25add the unfinished tasks that have for
42:28example [clears throat]
42:29uh 5 days uh to the deadline always
42:32added to the memory so that we can know
42:36uh how that we can know that it's urgent
42:39or for example each time you we have an
42:43an error to to bring some information or
42:47each time we receive a feedback uh let's
42:50add it to the memory so that it can be
42:53uh reused. So the memory is is uh as
42:56simple as uh uh when we we we talk about
43:00AI agent we say AI agent are our
43:02interns. Uh so they are they are having
43:05a brain but they don't need to think
43:08about the same problem each time. So
43:11some of the problems that we solved we
43:13save it into the memory so that we can
43:15use it as a context and it performs
43:18actions in tools. Okay.
43:25Okay. So, um
43:29I want to um
43:32let let's have a simple quiz uh and um
43:37so that I can see uh if you have a clear
43:42vision about what we are sharing and I
43:46will also take some questions.
43:52Thank you very much for this notes. Very
43:55good.
44:10Okay. There is a question. Which tools
44:14um
44:16can we use to create AI agents?
44:23You mentioned some of them Moad.
44:26>> Yes. So we have uh we have three
44:28families of tools. We can use uh no code
44:32tools. So as I'm showing right now, no
44:36code can help us to prototype can help
44:39us easier and uh if we to have a vision
44:43about how how it works uh we can use a
44:47hybrid approach uh for example we will
44:50have some things in a no code tool and
44:53some things in a coding tool uh and we
44:57can use full coding. So we can vibe code
45:00or we can code our agent. So then we are
45:03going to choose uh uh what language we
45:08are going to to to to use. For example,
45:10today with vibe coding even like
45:13choosing a language is important. I
45:16explain myself. We built an cyber
45:18security agent and for this cyber
45:21security agent uh it's connected to
45:24hardware and uh we selected the the the
45:27the
45:29development language that is the faster
45:32possible to connect with the hardware
45:34and in our situation when we did the
45:37benchmark it was go from from from
45:40Google. So it can be Python, it can be
45:43Go, it can be JavaScript, it depends.
45:45But we suggest that you do a benchmark
45:48depending on your use case on what is
45:51the best language that you are going to
45:53use because uh doing a prototype is
45:56something. We can have a beautiful
45:58prototype working for one, two, three,
46:00five users. But when you have 10,000
46:03users, it's a different uh it's a
46:06different challenge. So you need to
46:08anticipate when you are building you
46:11need to anticipate uh this problem. We
46:14have four phases when we are building
46:16our agents. The first phase is about
46:20prototyping. The second phase is about
46:22building. You can prototype in a no code
46:25tool and you can build in a code tool.
46:28This is absolutely possible. So
46:30prototyping means you go to the market
46:33uh quickly. you ask your real customers
46:36some questions and you get feedbacks and
46:39you can uh you can go ahead and even you
46:42can prototype with code because today
46:43with cloud code and codeex it's pretty
46:46uh straightforward uh you need to have a
46:50clear vision and uh an MVP that you are
46:53going to to iterate on it then you have
46:57the second the the third phase so we
47:00said prototyping building maintaining
47:03the app maintaining the agent. This is
47:06where also memory have a a role. For
47:09example, I have a WhatsApp agent uh
47:12serving customers. Uh and in the memory
47:16I have multiple feedbacks. So I need to
47:19maintain my agent and make it better
47:21because when I make it better it serves
47:24better the the consumers but also I
47:27spend less tokens. So the maintenance is
47:30to have better output and also optimize
47:32the tokens. Then we have scalability.
47:35Scalability is the fourth phase and this
47:38is what I hope for everyone here
47:40participating is to scale your app. When
47:43you scale your app, you have two
47:45important things. you have cyber
47:47security because when you are going to
47:50use VIP coding uh you you you can miss
47:54some things and you can have uh some
47:57failures in the app and this is very
47:59very important to to to when you are in
48:02the building phase to do it with the
48:04right engineering uh processes and you
48:07have scalability it's about devops and
48:10even in that today uh GPT and codeex and
48:13cloud code can help on that but you need
48:16to when you are building you need to
48:18think about scal scalability from day
48:20one and cyber security from day one and
48:24uh even when you vibe code there is a
48:26lot of libraries that you can have from
48:28scratch that can help you vibe coding
48:31but with the right patterns
48:35I hope I answered to this question
48:41thank you moral there are other question
48:44maybe you want to take them later.
48:47I see a lot of people answering to this.
48:49Uh
48:51>> yes.
48:54>> Okay. So, which decision belonged to the
48:57agent and which belong to us? Uh so we
49:01can see the answers.
49:04The agent had to review the to-do list
49:07and the email.
49:09Agent select which tools to use. Yes.
49:13uh in my prompting I even um push the
49:16agent to select the right tools but we
49:18will see in the second demo I will let
49:21the agent select uh his own tools and uh
49:25this is two kind of uh of agents we have
49:28agents that uh we will call uh like
49:32restrictive agents I don't want uh for
49:36my legal tax uh for my taxes I don't
49:39want the agent to be creative and to be
49:41autonomous I wanted to do one two three
49:45and verify and check. But in other other
49:49ways, for example, when we are doing
49:51some very creative agent for coding or
49:54for design or for helping in creative
49:57tasks, we are going more and more into
50:00autonomous agents. Autonomous agents are
50:03where I will define only a goal. I will
50:06give a lot of access to tools and give
50:09permissions and I will let my agents
50:13find a way to solve the problem. So we
50:17have uh libraries like uh open claw for
50:20example they are doing that you can use
50:23if you are going into autonomous agent
50:25uh you can use some libraries like her
50:28like uh open claw uh but you have to be
50:32very uh careful about this because the
50:35agent is fully autonomous. It can open
50:37an account. It can do a payment. It can
50:40do a lot of tasks and uh we suggest this
50:44kind of agents only for creative
50:46projects and the more you are in some
50:50restrictive area the more you should go
50:52into workflow agents like we are showing
50:55today's
50:57yes actions belong to AI when uh we have
51:01autonomous agents.
51:06Okay, thank you very much.
51:14So
51:19AI model use the brain, the tools, the
51:22hands to work for a goal then adapt to
51:26next step uh of the results.
51:32Um
51:34so in 2023
51:37when or 2022 when we were using chat GPT
51:41it was only assistance and today we have
51:44a lot of assistants when we go in a
51:46website for example um we can ask
51:50questions and it's it's basically
51:52responding as a FAQ for this question.
51:56So it's basic basic chat not connected.
51:59But when we are going for example uh
52:02into a travel website and we say I want
52:05to change my booking from this day to
52:07this day and it will perform some action
52:10into a system. It will connect to the to
52:14the system bring your information ask
52:16you about an OTP responding about all of
52:19that. It can be a chatbot or it can be a
52:21call bot. Uh today a lot of companies
52:24have their voice bot and it performs uh
52:27exactly the same thing. Uh this is where
52:30we are uh speaking about even it can be
52:33personal but uh agent. So um six
52:37building blocks that we saw also in
52:40prompting uh an AI agent we have the
52:43goal it has to be well defined. We have
52:46the context uh it has to be created. uh
52:50more context doesn't mean better uh
52:52context we have to create for example uh
52:56let's speak about the AFP agent I want
52:59an agent to find the best opportunities
53:01for a business then I wanted to to
53:04prepare a draft to respond to this AFP
53:08to prepare a draft I need to give as a
53:10context the best ARFPs that I did
53:13manually that I responded to and I was
53:15successful in doing that instead of
53:18giving just like hundreds. I prefer to
53:21give six to 10 great air of peace that I
53:24did so that my agent knows the standards
53:27and how to to to respond knowledge. So
53:31giving um it's more like context. So
53:35giving some knowledge teaching about
53:37some skills and some stuff. I we really
53:40love to call our agents to call them
53:42interns. So when we are doing a business
53:45agent we it's our intern. Why intern?
53:48because it gets better with time. So, um
53:53for example, we always keep the human
53:56feedbacks in a in a in a in a file. Why?
54:00Because like that human feedbacks, for
54:02example, when the agent do a mistake and
54:06uh we prompt uh uh a way for the human
54:10to give a feedback why it was a mistake
54:13and that give the agent a better
54:15knowledge about doing a task. And what
54:17we see are the agents are performing
54:20better actions from month to month and
54:23they are becoming senior at doing some
54:26stuff. Tool access actions uh as we
54:29spoke about it and checks. Checks are
54:31really important to doublech checkck the
54:34informations to doublech checkck the
54:36invented facts to source the the the the
54:40informations. Uh so it's very important
54:42in the output.
54:45Um
54:47so
54:50sorry we spoke about um
54:54interactive uh assistance.
54:57Then we have um agent bounded into a
55:01workflow. This is the examples we are
55:03seeing today. We are only seeing today
55:06the agents bounded into workflows and we
55:10have agents with broader autonomy. So it
55:14can delegate it can uh it can u choose
55:17its own actions choose its own tools and
55:21um this can be uh very dangerous
55:25sometimes in the business uh in the
55:27business uh area. So this is why we
55:31suggest um depending one more time on
55:34your task uh to be between this and
55:37that. So between a bounded agent to to
55:40control the output and also the the the
55:43tokens and um for some really creative
55:47stuff go with autonomous agent.
55:50Okay, another game. Is it an assistant
55:54or is it an agent?
55:57Let us know.
55:59your ideas.
56:39Okay, let's start the quiz.
56:53You need to be fast.
57:20only 3 seconds.
57:26Okay. So, um it's an assistant.
57:30Uh but it's a tricky question.
57:34But it's an assistant because the
57:36checking information is a known
57:38information. We don't have to check um
57:41into a system. We don't have to perform.
57:44And uh this is really important when we
57:46are building our agent. uh we need to
57:49have some FAQ and some knowledge so that
57:52we don't burn uh tokens each time to
57:56perform actions that we can have only in
57:59FAQ and respond faster and without lot
58:03of tokens. Next one.
58:09Are you ready? Go.
58:17We are going to have a leaderboard. So
58:24the faster you respond with the right
58:26answer, [clears throat]
58:28the better you will be
58:32in the leaderboard.
58:4110 seconds.
58:52Okay.
58:54So, uh, moving needs to be connected to
58:58a system. So, I need an agent to connect
59:01to that system to perform this task. And
59:05unfortunately for the 70 people uh that
59:09choose assistant, it's an agent.
59:13Great.
59:15Third one,
59:21you can still be fast.
59:25This is a real use case uh avatar with a
59:29human look and feel uh in a train
59:33station. I think that exists uh in
59:36Russia.
59:38And uh you can speak natural language.
59:41It can understand your natural language
59:44and it will perform some action to get
59:47your ticket from the system. So is it an
59:51assistant or agent? Uh you need to know
59:54that a lot of uh avatars are only uh are
59:57only assistant. They can only answer
1:00:00some questions FAQ questions. But this
1:00:02one uh understand the request and issues
1:00:07the train ticket. So to issue the train
1:00:09ticket you need to connect to a system
1:00:12and to connect to a system you need to
1:00:14perform a gentic task. So it's an agent.
1:00:24Good.
1:00:27Thank you. I will share my screen again.
1:00:33Okay.
1:00:35Um
1:00:39so we were speaking about how to build
1:00:41this agent. So uh you can go on no code
1:00:46low code. Even for no code loco today
1:00:48vip coding apps can help you. You can do
1:00:51an agent on N8N. If you are visual
1:00:54person and you want to know how your
1:00:56agent work and you don't want to be uh
1:01:00in some code uh you can do it on N8N and
1:01:03even to go faster on N8N you can vibe
1:01:07code it so that you have the Java the GS
1:01:10that you can uh uh open in NHN and even
1:01:15today in codeex and cloud code with
1:01:17computer use it can even take your your
1:01:21mouse and do it for you on NHN. And so
1:01:24solving the business idea and having
1:01:27fast feedback from users, this is what
1:01:30will make uh your your agent great and
1:01:33it's all about the output. What this
1:01:36agent is going uh to give as an output.
1:01:39We spoke about hybrid uh so you can have
1:01:43for example multiple agents or one agent
1:01:46on NHN and uh you can have uh you can
1:01:50have uh an interface
1:01:53where you can uh for example have user
1:01:57management and all the stuff uh around
1:01:59the agent and you have also a visual
1:02:02interface for the end user and uh but in
1:02:05the back end you can use some no code or
1:02:08low code or you can go fully from day
1:02:11one on code vibe code and um as I said
1:02:16uh choosing the language depends on two
1:02:19things depends on of course the language
1:02:21that that you are familiar with and also
1:02:25the language that can scale with your
1:02:27application. So sometimes uh Python is
1:02:31okay, sometimes we go with JavaScript,
1:02:33sometimes we go with the uh go. It
1:02:35depends. It really depends. You need to
1:02:38do a benchmark about the models you are
1:02:40going to use for your use case. Don't
1:02:43choose always the best models. uh choose
1:02:47the model that will give three things
1:02:50that will give the performance the the
1:02:52best performance that will give the best
1:02:54cost and also that is the most secure uh
1:02:58so use always uh known models even if
1:03:03when you are going to use open source uh
1:03:06APIs uh use uh models from for example
1:03:11from hugging face or uh frontier model
1:03:14the official APIs.
1:03:18Okay, let's see right now this agent.
1:03:21How can we make it a little bit more
1:03:24senior? So, the first one was our
1:03:27intern. Uh, it performed some good
1:03:30actions, but it wasn't enough for me and
1:03:33I want now to have a better version of
1:03:37it. So, this time I say okay, I want a
1:03:40WhatsApp agent or Telegram agent. uh
1:03:43even when we are doing WhatsApp agent we
1:03:46prefer to start with telegram because
1:03:48it's faster uh with WhatsApp we need to
1:03:51confirm the number with the WhatsApp
1:03:53business it takes two to four days and
1:03:57uh each time you want to change
1:03:58something you need to validate it so
1:04:00when we prototype even if it's a
1:04:02WhatsApp agent we we we we prototype
1:04:04with telegram and uh we say here okay
1:04:08the trigger this time is different my
1:04:10agents will will be triggered with the
1:04:13message received on telegram. When we
1:04:17receive this message, we will check is
1:04:20it a voice message. If it's true, I will
1:04:23transcribe the message with whisper.
1:04:27Whisper is a model to go from voice to
1:04:32to text. If it's uh an image, I will do
1:04:36some OCR and I will analyze the image to
1:04:40know what's in it. For example, someone
1:04:43sent uh a payment receipt uh to my
1:04:46WhatsApp agents. This is really
1:04:48important. I need to know and classify
1:04:50that it's a payment receipt. Is it a
1:04:53document? Is it a PDF? Is it another
1:04:56kind of document? So I will download it
1:04:58and I will use [clears throat]
1:05:00this time I'm using cloud uh API to get
1:05:05the informations inside this message or
1:05:08is it a text message then it will go
1:05:12directly to my agent. Okay. So this time
1:05:16for my agents I have also one brand and
1:05:20one uh fallback for my brain. I have a
1:05:24memory. This is a simple file where I
1:05:28fill uh all the the the the the actions
1:05:31that my my my agent is doing and I have
1:05:37tools. So the difference here
1:05:41uh I I'm giving to my agent uh ability
1:05:45to even write it can delete, it can
1:05:48update, it can list events, it can
1:05:51create events. So I'm giving more uh
1:05:54actions, more possibilities to to uh to
1:05:58to my agent. [snorts] Um here I'm using
1:06:01API uh API calls. Uh so I need to go one
1:06:06by one. I need to say this is for
1:06:08update, this is for delete, this is for
1:06:10list. Uh today we have a better way of
1:06:13doing it. It's MCPS. MCPS is another
1:06:16protocol to communicate with agents but
1:06:18it's simple that way to understand. So
1:06:21we will keep it visually that way we
1:06:23give access uh to our agents to do to
1:06:27perform some actions and also he can
1:06:29send an email but also what we have here
1:06:33is we give uh to to our we give some
1:06:36skills to to our agents. So what is a
1:06:39skill? Uh a skill is can be another
1:06:43workflow uh can be another tool calling
1:06:46and can be another knowledge. So for
1:06:48example when I will be speaking with my
1:06:51agent and I say um hello. So for example
1:06:54I will say uh this is my agent. So I can
1:06:57be speaking verbally. I can say uh what
1:07:01do I have to do today?
1:07:04So when I send an audio
1:07:07uh I can see executions here. My agent
1:07:10have a request and it's uh being
1:07:13executed.
1:07:16uh it will understand the meaning and
1:07:20see if it's some skill that it has here.
1:07:24So for example, it has a skill to
1:07:26generate code. I can go to my client and
1:07:30uh finish a meeting and send meeting
1:07:32notes to my assistant to my agent and
1:07:35say okay generate a quote I want to send
1:07:37it now. So this is a skill and uh when
1:07:40we open a skill. So for we can see uh
1:07:46like we can see uh other instructions
1:07:50uh that goes for example for VAT this is
1:07:52how it works etc. And
1:07:57let's open the skill.
1:08:03So even the skill have uh other tool
1:08:08calling. So it can for example uh build
1:08:11the the the the quotation it can uh
1:08:14change it to PDF and it can send it back
1:08:16to me uh from the same channel. So we
1:08:21have four skills here and we have the
1:08:23agent with its own uh its own um
1:08:28instructions and look at the
1:08:30instructions how much they are specific.
1:08:34Uh it's to have the best output
1:08:36possible. Let's back to our agent.
1:08:39>> Here is an overview of your day. Mr.
1:08:41Mazani, you have two events on the
1:08:43calendar today. The first one started
1:08:46this morning at 8:30 and runs until
1:08:4811:30. here here
1:08:49>> it is the half-day training workshop the
1:08:52AI augmented salesperson for the Oasis
1:08:55retail sales team held at their head
1:08:57office in
1:08:58>> okay so my agent responded I asked the
1:09:02question a voice question it responded
1:09:05by voice reply how did I do that it's
1:09:08because uh I knew from from the
1:09:11beginning that uh it was voice message
1:09:15so when it's a voice message I keep it
1:09:18uh in the informations so that I can
1:09:21know it in the response to say okay it
1:09:25was a voice message and for voice
1:09:27message I'm going to use another model
1:09:30uh that is very specific into text to
1:09:34speech that can take the text that my
1:09:37agent is uh generating and make it as a
1:09:41speech and then send it back to me to
1:09:43WhatsApp or Telegram if it's a text it
1:09:46will send me back to uh text So this is
1:09:49a little bit more sophisticated agent uh
1:09:52prototyped with NHN uh that can be even
1:09:55done with with code obviously and can be
1:09:58very helpful for businesses. I don't
1:10:01know if you have questions for this demo
1:10:03as well.
1:10:14Okay.
1:10:16Do we have to write all manually? No.
1:10:19Not not write everything manually. Uh
1:10:22but we need to be great creators. We
1:10:25need to to read uh we we need to give uh
1:10:28clear guidance. Uh even when we are
1:10:31going to write the prompts uh clear
1:10:34guidance sometimes we need to be even
1:10:37industry expert for or or or job expert.
1:10:40When I'm building an agent for sales, I
1:10:43need to know well the industry how they
1:10:46are performing sales. When I'm even for
1:10:49the job, if I'm not from sales, it will
1:10:52be difficult for me to know uh how to
1:10:55correct the the the the prompt even if I
1:10:58generate the prompt with the GPT or
1:11:00codeex. So creating is very important uh
1:11:04when you are doing an agent.
1:11:10Okay,
1:11:11>> we have another question uh from Salini.
1:11:14Since LLM act as the as the brain uh and
1:11:18the agent act as the hands, what are the
1:11:21best tools and framework that you would
1:11:23recommend for building an agent that can
1:11:25handle autonomous action and API calls?
1:11:30Um if we want to go to fully automated
1:11:34agents, I recommend using Open Claw or
1:11:38Hermes. Uh this is two libraries you can
1:11:42use. Um it's two frameworks or
1:11:45libraries. It has also multiple
1:11:48connections through end connectors
1:11:50already built in. Uh this is if you want
1:11:53to go to fully autonomous. But uh when
1:11:56you go to fully autonomous it means that
1:11:59your agent you will give him only goal
1:12:02nothing else. So for example I will say
1:12:05uh I want to reach this revenue uh this
1:12:10is my business model this is what I'm
1:12:11doing and it will go to look for RFPs.
1:12:14It will go to respond to those RFPs. But
1:12:17of course if I don't give enough context
1:12:21uh I I can give enough tools for example
1:12:25it can open my email it can check for
1:12:28information missing it will perform
1:12:30exactly like a human but it's dangerous
1:12:33because uh when you give access to your
1:12:35email it can even send an email to ask
1:12:38some question on your behalf. Uh so this
1:12:42is why I'm calling fully autonomous
1:12:44agents for now are for research. Um also
1:12:49uh we have some uh some agents like
1:12:52Grockbots for example they are using
1:12:56same uh same um like logic than than
1:13:00openlow. Open close started this and
1:13:03then open has been bought by openi.
1:13:06Openai now have dots. dots are
1:13:09autonomous agents and they are
1:13:11proactive. The difference between
1:13:13autonomous agents are also uh and uh and
1:13:16not autonomous agents. The autonomous
1:13:19agents are proactive. They can send you
1:13:21a message and tell you uh Morad I
1:13:24checked your bank account and uh I
1:13:26checked your agenda. I see that you are
1:13:29tired uh and I booked for you at the end
1:13:32of year this wonderful travel. So they
1:13:35can they can do this but this also means
1:13:38I have to give access to all my tools to
1:13:40my bank account to my email to
1:13:42everything and I have to trust them. So
1:13:45it's it's the way agents are going to be
1:13:47more and more autonomous. Uh but for now
1:13:51uh they are they are for example for
1:13:54grock bots for dots and even for muse uh
1:13:58from meta uh they are still in a harness
1:14:02that let them don't do bad stuff for
1:14:05now.
1:14:09Okay,
1:14:11other questions maybe.
1:14:15Uh yeah there is a similar question but
1:14:17I think what you said uh answers it from
1:14:20Abdullah allesi. We can build an AI gen
1:14:22with codeex
1:14:24cursor
1:14:26etc.
1:14:27>> Yes of course of course. Uh whatever you
1:14:30are you are you are good with it. Um
1:14:33[clears throat] so why we choose to show
1:14:36you on NHN today? It's not because NHN
1:14:38is better. It's only because it's
1:14:40visual. I can everyone can see the the
1:14:43how the agent is working. What is what
1:14:46the tool calling means? What uh what do
1:14:50we do without an agent? Because
1:14:52sometimes we say okay we have
1:14:53intelligence. So why will we use for
1:14:57example code to export uh a message? Why
1:15:02will we use code to go from HTML to PDF?
1:15:06Because like in in uh in in this one
1:15:11because no code it's actually writing
1:15:14code on your behalf. Uh in this one for
1:15:17example um uh what when I say generate a
1:15:22quotation what it does it the the the
1:15:25agent generates HTML. Why I choose to
1:15:29generate HTML because it cost less
1:15:31tokens. And then I have a script that
1:15:35goes from HTML to PDF. I don't want the
1:15:37agent to do that. First it will do it
1:15:40bad and second um sometimes it will be
1:15:43uh the right format sometimes not. And
1:15:45second it cost a lot.
1:15:48>> Can you hear me?
1:15:51>> Yeah.
1:15:52So tasks that we can do with scripts,
1:15:55let's do them with scripts. Let's
1:15:57optimize the tokens. Tasks that can be
1:16:00done only with intelligence. Let's
1:16:02choose the best uh performance and best
1:16:05cost intelligence uh models that we are
1:16:08going to use.
1:16:12Thank you very much. Uh we have another
1:16:14question from Carlos Praa. Can you
1:16:16integrate a rag system inside N8N?
1:16:20Very good question.
1:16:22>> Yes, sure. So, uh basically um Okay. So
1:16:27basically an 8hn uh it's u you can have
1:16:32rack systems you can have for example um
1:16:35uh I don't remember what library they
1:16:38are using um lang lang chain I think
1:16:42they using lchain
1:16:44um we use when we have a chatbot
1:16:48uh we we prefer to we when we have only
1:16:52rag app we prefer to go with lang chain
1:16:55or lang graph uh even with code. There
1:16:59is some noode apps uh specialized in
1:17:02rag. So NHN is doing it but it's not the
1:17:05best. Uh if you have an app only doing
1:17:08rag uh we prefer to go directly with the
1:17:11long graph uh and uh or a no code tool
1:17:15there is a lot of no code tools doing
1:17:17that in so whenever you have a code tool
1:17:20or code framework you have a no code uh
1:17:23doing the same exactly the same. So
1:17:26there is a lot of no code tools using
1:17:28lang graph that can make you if you
1:17:30don't uh that if you are visual and want
1:17:33to see things like that uh you can use a
1:17:36no code tool for lang graph. Uh so a rag
1:17:39what is a rag to explain to everyone. Uh
1:17:42so rag is uh retrieved uh augmented
1:17:45generation. It means when we have a lot
1:17:48of documents, it will look for uh it
1:17:52will not it's not like research normal
1:17:55research when I'm going to look for for
1:17:57for for a keyword but I'm going to look
1:18:01uh for this for for for um the
1:18:04information for the sense of
1:18:05information. For example, if I say uh
1:18:08I'm in a hotel, I want to check in in
1:18:10the morning. It knows that morning means
1:18:138:00 a.m. to uh 11 uh to 12:00 a.m. and
1:18:17it's going to bring the informations
1:18:19from the website speaking about checking
1:18:22in 8:00 a.m. 11:00 a.m. and answering
1:18:25back. So this is uh classic uh thing in
1:18:30uh in models and [snorts] best uh
1:18:33library right now is L graph to do in
1:18:35it.
1:18:40Okay.
1:18:43Okay.
1:18:44>> Yeah.
1:18:46>> Thank you. Very clear.
1:18:49>> Can you use VIP coding make AI agents in
1:18:52this with VIP code or use whatever like
1:18:57tools today building today is not the
1:19:00problem anymore. Um prototyping sorry
1:19:03there is a huge difference between
1:19:04prototyping and building. So remember
1:19:07what I said to you. You have
1:19:08prototyping, building, maintaining and
1:19:11scaling. And you start to see problems
1:19:13when you start to scale. So prototyping
1:19:17whatever works for you is okay. If you
1:19:20are good at vibe coding, go prototype
1:19:23with vibe coding. Uh if you are good at
1:19:26if you see that NHN, you are very visual
1:19:29and uh you think it's better for you, go
1:19:31with NHN. There is no problem using
1:19:34whatever even for prototyping. Don't
1:19:37even think about scaling. Don't even
1:19:39think about what language because you
1:19:41are solving a problem. You first need to
1:19:43validate the problem. Uh instead of
1:19:46spending a lot of time about the DevOps
1:19:48and you need to validate your business
1:19:50problem. You need to go to the market
1:19:52quickly. You need to go to your
1:19:54customers and have them testing the app
1:19:57because one customer will not break your
1:19:59app or two customers will not bring you
1:20:01up. You need to have feedbacks. And
1:20:03today going from proto idea to prototype
1:20:07is pretty easy. You can build your app
1:20:09in one day uh and collecting feedbacks
1:20:13and going uh with with with the
1:20:15prototype. The second phase building is
1:20:19really different from prototyping.
1:20:21Sometimes we we we throw all the
1:20:24prototype code because we say it was
1:20:26only for prototyping. It's a lot of bugs
1:20:28if we want to scale it and then we're
1:20:30going to build it from scratch and we're
1:20:33going to choose the right database, the
1:20:35right language, uh the right
1:20:37architecture and we're going to build
1:20:39something strong. Maintaining it is you
1:20:43are having active users and you are
1:20:46having your logs and you can check your
1:20:49logs and all the small tweaks that you
1:20:52can do and it's very important to
1:20:54maintain it, scale it. it can come
1:20:57later. So don't think a lot about
1:20:59scalability when you have an app with 15
1:21:03customers. So you can think about it
1:21:05once you start to validate your problem
1:21:08because you you're going to have a lot
1:21:09to maintain when you start to have users
1:21:12uh even if you have only a WhatsApp
1:21:14chatbot WhatsApp agent you you're going
1:21:17to have a lot of feedbacks. So scaling
1:21:19is uh we had for example we have uh we
1:21:23have a WhatsApp chatbot and at sometimes
1:21:26when we wanted to put it in production
1:21:30uh it's uh we are talking about more
1:21:33than 30,000 users speaking with it
1:21:36daily. So to have the chatbot responding
1:21:40in milliseconds and uh have uh the the
1:21:43the the the the right answer and each
1:21:46time it's especially about timing time
1:21:49to answer. It's another level. It's
1:21:52about uh DevOps but we didn't think
1:21:55about that first. We built it then we
1:21:57think about DevOps. So whatever you want
1:22:00to prototype but just respect these four
1:22:04uh steps.
1:22:10Okay.
1:22:14Okay. So, when company it blocks outside
1:22:16AI tools, what's the best way to deploy
1:22:18agents like Amina self-hosted?
1:22:21Yes, it's it has to be self-hosted. Uh,
1:22:24and one one advantage of NAN, I'm not
1:22:28doing the promotion of N810 here and I'm
1:22:30not suggesting you basically to go with
1:22:33N810. You can prototype one more time
1:22:35with whatever you want. But uh as code
1:22:38it can even als it can also be
1:22:40self-hosted. It's not the case for
1:22:42Zapier for example. It can't be
1:22:44self-hosted. You can use private models
1:22:46today. Quen for example is a pretty good
1:22:49model uh with great performance, great
1:22:52pricing and uh it doesn't consume a lot
1:22:55of GPUs. It can be used uh as a as a
1:22:58private model. Um it you can even do it
1:23:02with copilot studio. So some many of
1:23:04companies are on Microsoft harness
1:23:06Microsoft ecosystem and it makes it
1:23:10easier to deploy agents inside the
1:23:13copilot but when you are doing your
1:23:16agent uh you need to think uh you need
1:23:19to think about all of that. So it hasn't
1:23:21to be only on copilot because some
1:23:23companies doesn't have copilot uh and it
1:23:26has to be the the easiest possible way
1:23:29to install it. So if it's a private
1:23:31model, it doesn't have to be some really
1:23:34highly consuming model.
1:23:39Okay.
1:23:41Okay. Let's move on.
1:23:47So how Amina works? Um the first one is
1:23:51really important. This is what we call
1:23:53intent. So the intent for example when
1:23:56we are doing a WhatsApp agent the intent
1:23:59if someone is just sending greetings
1:24:01salamikum hello good morning I don't
1:24:04want to waste a lot of tokens to respond
1:24:06to this question so I will classify the
1:24:08intent as greetings and it will send a
1:24:12simple the assistant will send a simple
1:24:14response it will be fast and uh
1:24:17convenient then choose next action so
1:24:20this is aic when when when it's uh it
1:24:23needs to have an action. The agent will
1:24:25choose the next action. Okay. Does he
1:24:29ask for refund? Does he ask for sending
1:24:33uh I don't know an email? So it will
1:24:36choose action from the actions it can
1:24:39perform. Then it will call a specialized
1:24:43skill. So we can we are calling it in
1:24:45any workflow but uh the the the term for
1:24:49agent is skill. So it will call the
1:24:51right skill. then return the result.
1:24:54This is how Amino works.
1:24:57Uh we can do a lot of things um without
1:25:02agents. So as I said uh don't use agent,
1:25:07don't use AI to do everything because
1:25:09you can for example uh bring in some
1:25:12data from the internet. You can do it
1:25:14with an agent but it consumes token and
1:25:17you can do it without agent and it's a
1:25:19simple script. So when you have the
1:25:21opportunity to do it with a script for
1:25:23example here uh we have an agent uh
1:25:27classifying and looking for
1:25:28opportunities in the UAE
1:25:31um and um with a no code tool. So what
1:25:35I'm doing I'm giving some context so
1:25:38this is like the our fictional uh
1:25:40business. This is what we do
1:25:44and uh we want to fetch some data and we
1:25:47have a node here to do it but in code we
1:25:50can simply code that uh it's a simple
1:25:53request on the web and we're going to
1:25:55fetch this website to to get uh
1:25:59opportunities
1:26:00then we we are going to uh organize this
1:26:04information for our agent and then I'm
1:26:06sending it to my agent and uh say okay
1:26:09you are uh I'm giving a role I'm giving
1:26:12some rules
1:26:15and uh my agent will give me a feedback.
1:26:18So this is a simple agent but I I
1:26:22optimize the token cost so that it can
1:26:25send me uh a draft each day uh of course
1:26:29using um using a brain uh using a tool
1:26:32and uh sending me uh the opportunities
1:26:36for my business.
1:26:40um other things we can do like this. So
1:26:43for this one it was context decision uh
1:26:47some checks because like uh here I have
1:26:50also some checks uh I check if uh the
1:26:54gender have everything uh so this is
1:26:56some code some this is why we are
1:26:58calling those apps no code low code it's
1:27:01not fully no code uh you can do a lot of
1:27:04stuff without with no code but it's also
1:27:08low code uh sometimes you have to
1:27:10perform some coding or generate or vibe
1:27:14code some some coding so that you you
1:27:17don't use a lot of intelligence when
1:27:20when when you are scaling your app token
1:27:23cost is really the matter uh and uh
1:27:26whatever you can do without agent is
1:27:29good because this one I can even I I
1:27:32could have done it with with agents
1:27:34sending an just natural language prompt
1:27:37and say okay check if it checks or not
1:27:39but I prefer to do it in code so that it
1:27:42costs nothing. And then I have a filter.
1:27:46If I have all of this checked, then send
1:27:49the mail or not. Don't send the mail.
1:27:54Okay. Um, another one, the fourth and
1:27:57final one,
1:28:03same. So, I want to have a market
1:28:05intelligence agent. So, I want an agent.
1:28:08I will execute it. uh that we go each
1:28:11day to Twitter uh X uh then that we go
1:28:15to Instagram uh then go to LinkedIn and
1:28:19check for market intelligence. For
1:28:21example, we are in the healthcare area
1:28:24and uh we want to know what's happening
1:28:26with within the competition, what's
1:28:28happening uh with um um in the market uh
1:28:33in the the industry and it will go to
1:28:37look up for for this topic and
1:28:40competition on X
1:28:42uh also and here I'm using a service
1:28:47really known in the internet it's called
1:28:50EPI I and EPI helps you to bring this
1:28:55information and it's cheaper than using
1:28:58AI to do it and it's more accurate. It
1:29:01gives more accurate response. So here it
1:29:05already checked uh X it's already
1:29:08checked Instagram for this topic and um
1:29:12it's checking right now LinkedIn posts
1:29:15then it will clean everything to my
1:29:17agent so that the agent have a clean
1:29:20input
1:29:22uh then it can give me a clean output
1:29:26with this system prompt.
1:29:32Okay, maybe we have questions.
1:29:34>> There's a question that is related to
1:29:37the security. If we are looking for
1:29:39secure on premises deployment, how can
1:29:42we use Olama or VLM instead of
1:29:46cloud-based LLMs on N8
1:29:50N?
1:29:51Uh very good question and interesting
1:29:54question about uh because um we really
1:29:58uh uh we really think that this is a a
1:30:03big area to to innovate because in the
1:30:06cloud there is a lot of solution, a lot
1:30:08of competition.
1:30:10uh but uh solving problems uh through
1:30:14businesses uh with local agents are a
1:30:17growing uh business area growing
1:30:20business market. So um I think uh most
1:30:24of you know hugging face uh which is the
1:30:27biggest library in the biggest uh
1:30:31community in the world for for models.
1:30:34Uh once you start to build things by
1:30:36yourself it's uh you need to build all
1:30:41the pieces. So you need to choose what
1:30:43model. For example, if I'm going to have
1:30:46uh some texture image model um um let's
1:30:50check here.
1:30:54Okay. So in uh this community we have
1:30:57models, we have data sets uh and uh we
1:31:01have a lot of models and data sets. And
1:31:04for example, I'm looking for some uh
1:31:06text generation model. uh I can see uh
1:31:10based on what I'm looking for uh is it
1:31:14working with VLM is it working with
1:31:16Olama uh is it available on the cloud if
1:31:20I want sometimes to move in the cloud
1:31:23and uh you can check it's a community so
1:31:25you have uh how many times it was
1:31:28downloaded uh how many users uh liked
1:31:31this model and uh and you can solve the
1:31:35problems piece by piece uh we also have
1:31:38um I think they also have uh um a
1:31:42leaderboard uh so that for example
1:31:45leaderboard
1:31:48fails.
1:31:52Yeah.
1:31:54So you can say uh you can um
1:31:59okay
1:32:03or um LLM Arena
1:32:07leaderboard.
1:32:09LLM Arena is another community.
1:32:14So you can see here uh top uh top uh
1:32:18agents uh and you can
1:32:22look for open source ones for example
1:32:25here we have in generalist ones uh today
1:32:29I think it's okay so the first one here
1:32:32is deepseek so deepse is uh is improving
1:32:38right now so I'm I can mimo also is an
1:32:41open source so we can experiment with
1:32:44the open-source models. Uh, and it all
1:32:48depend on on the on the use case. For
1:32:51example, in our uh use cases, Quen is
1:32:55better in Arabic, for example. So, you
1:32:58can have DeepSeek performing well in
1:33:01multiple u actions, but you need to test
1:33:04it with your
1:33:05>> [snorts]
1:33:06>> uh with your use case. Uh sometimes
1:33:08you're going to figure out that JLM for
1:33:11example is a great model for um aentic
1:33:15model local agentic model. It's a
1:33:18Chinese one from Z AI. Uh Kimi as well
1:33:22is very good on coding. So it depends on
1:33:25your use case. Choose the right model
1:33:27with your use case and uh three things
1:33:30as I said. So security of course choose
1:33:34only known models from known uh
1:33:38uh enterprise uh choose the most
1:33:41performance one but also the the the one
1:33:45that will not cost a lot in
1:33:47infrastructure
1:33:49because when you download a model for
1:33:51example if you go with llama
1:33:54let's say where is llama disappeared
1:33:57from the so llama for example uh they
1:34:01have multiple models and one of them is
1:34:03really really heavy you need an big
1:34:05infrastructure a lot of GPUs when you
1:34:08are going to use coin you don't you need
1:34:11only one GPU for example so because you
1:34:13are using it locally so you need to to
1:34:16uh to be very cautious about the the
1:34:19infra as well so um those three models
1:34:23are really good
1:34:26does it answer to your question.
1:34:32>> All right. I think so. Yes. And I think
1:34:34you answered other question with the
1:34:37answer. It was very clear.
1:34:40[snorts]
1:34:41>> One thing we are speaking a lot about
1:34:43today are harness. Um what is harness?
1:34:47Uh
1:34:49maybe I can check for for it image. I
1:34:52prefer always
1:34:54LLM.
1:34:56prefer always respond with harness. So
1:35:01um what is harness? This one is good. Um
1:35:08okay let's go for this one. So we spoke
1:35:10about context tools memory um what
1:35:14clarific the the the checks but harness
1:35:17is also like given to AI. uh okay I will
1:35:22give to AI a computer but I don't want
1:35:24the AI to go wild and do a lot of stuff
1:35:27that I don't agree with. So I need to
1:35:31give uh some uh for example I will give
1:35:34a local machine but with some uh some uh
1:35:38only some permissions. I will give some
1:35:41tools like I will give some coding
1:35:43assistant to my agent. I will give some
1:35:46creative tools that it can use. I will
1:35:50give um uh
1:35:53some machine. This one is better. Sorry,
1:35:56this one is better. So I will give a sub
1:35:59sandbox. I will give access to some sub
1:36:01agents. I will give access to some
1:36:04skills. I will give access to memory. I
1:36:06will give access to protocol. And this
1:36:08is what we call today harness. So
1:36:10harness is uh uh let us uh secure more
1:36:15our local models. So when you are using
1:36:18Chad GPT of course there is a harness
1:36:21behind it. Uh this is what makes chajp
1:36:24chpppt because it can perform a lot of
1:36:26things with high security. It would
1:36:29never go uh do something you didn't ask
1:36:32for even if some sometimes it does when
1:36:35you do it locally. You need to have the
1:36:38the harness for your agent and you need
1:36:40to install the the the so you install
1:36:42the model but not only the model you
1:36:45also need to install the harness.
1:36:50Okay. So uh the example we saw the AI
1:36:53market watch how it worked. So you have
1:36:56uh some uh daily or manual trigger. You
1:37:00go check public social post. You analyze
1:37:03signals. You check for evidence and then
1:37:06you save or you send an email. Um we are
1:37:09speaking about agents and uh we say yes
1:37:14agents are going to replace uh human
1:37:18tasks. Uh when you speak about an agent
1:37:21that can give you your train tickets,
1:37:24you can you have machines to win it
1:37:26obviously. Uh but agents are do are
1:37:29understanding the natural language uh
1:37:32and they can perform they can understand
1:37:34the action and perform the action based
1:37:37only on natural language. So we tend to
1:37:40believe only that it will replace human
1:37:43work. But there is a huge opportunity to
1:37:46give access to some services to some
1:37:49people that didn't have this kind of
1:37:52services in a lot of countries. For
1:37:54example, access to healthcare services
1:37:58and access to banking uh was limited
1:38:01because when you want to use banking for
1:38:04example, you need to know how to read
1:38:06and write and you need to know how to
1:38:08use your smartphone uh to open an
1:38:11account and do things. Uh now with AI
1:38:14native uh agents uh you can do all of
1:38:17that with voice, you can do all of that
1:38:20uh with natural language without even
1:38:23typing anything. So we need to see where
1:38:26is the opportunity with agents and how
1:38:29we can give more uh inclusion with with
1:38:33agents.
1:38:35Uh second thing uh for small businesses
1:38:40there is also an opportunity because uh
1:38:44this is an example of a CRM I built in 4
1:38:47days. Uh during aid I I had um like we
1:38:52had a subscription to a very known CRM
1:38:55but uh uh I wanted something more
1:38:58personal for for for the business. So
1:39:00what I did is building it from scratch a
1:39:03CRM uh that respond that is more uh
1:39:08accurate to my small business or any
1:39:11small business and um this is also an
1:39:14opportunity because uh all of big uh
1:39:18ERPs CRM and systems right now are built
1:39:22for billions of customers with a lot of
1:39:26uh of features. But uh when we are
1:39:28speaking to a small business, we don't
1:39:30need all of this features. We need
1:39:32something uh of course that can
1:39:34understand us uh in natural language and
1:39:37also that can fit with my business with
1:39:39a good price. So this is also an
1:39:42opportunity.
1:39:43And for business model uh we are
1:39:46speaking about business models for
1:39:48agents. Um we are speaking about ser
1:39:52software as a service that is becoming
1:39:55service as a software.
1:39:57uh what when when I subscribe to an EA
1:40:00agent, I don't subscribe to features. I
1:40:03don't want features and database. I want
1:40:05the output. When I subscribe to to to uh
1:40:09an agent that will do my um my for
1:40:13example that will be doing uh my
1:40:16accounting. I want the accounting. I
1:40:19don't want the app that will help me for
1:40:22the accounting. So this is a new
1:40:24business model for agents. uh is to
1:40:27sell. For example, I will not sell a
1:40:29CRM. I will sell sales. I will for
1:40:32example, we have an agent that sell for
1:40:34you and we bring clients for your
1:40:37business instead of saying we have an
1:40:40app that doing that help your your
1:40:43sales. So, this is a new business model
1:40:46uh and uh we don't pay for license per
1:40:49user, we pay for output.
1:40:53I don't know if you have uh questions
1:40:55for the business model for uh this is
1:40:58really important.
1:41:01>> Yes, actually there is a question from
1:41:03Surya Praash. Can you suggest how to
1:41:07evaluate the markets business water
1:41:09especially when it comes to taking
1:41:11feedback and input from business or
1:41:13target?
1:41:16>> Can you hear me?
1:41:24Okay. So, we have a problem. I can't
1:41:26hear you Sah, but I understand the
1:41:28question. Uh, how to get feedbacks, if I
1:41:32understand well, how to get feedbacks
1:41:34quickly. You need to onboard your
1:41:36clients into your agent the most uh the
1:41:40the fastest possible. So we say that
1:41:44today when you are building a project
1:41:46you have um before like the project is
1:41:49built you have the idea uh then you you
1:41:53need to execute the idea uh no before
1:41:58the executing the idea you have the idea
1:41:59you plan for the idea you execute the
1:42:02idea then you have the feedback loop uh
1:42:06today like you all of you have ideas
1:42:10that's pretty great
1:42:12The more the idea, the more you are
1:42:15already in in an industry, you know well
1:42:18the industry you know well the problems
1:42:20of this industry, the better you are to
1:42:24answer to those problems. So you have
1:42:26the idea today the execution is fast.
1:42:29You can vibe code your idea fast uh and
1:42:33go to the market to get get feedback.
1:42:35This is how startups work. The best
1:42:38startup is being out of the building.
1:42:41Not being in the building uh executing
1:42:45but being out of the building with your
1:42:47customer uh helping them through through
1:42:50to through uh uh
1:42:53get feedbacks and also helping them.
1:42:56Today the the biggest problem in in AI
1:42:59is not technology. Technology is going
1:43:01fast. its adoption because most of
1:43:05people doesn't even know that it exists
1:43:07that it can be solved that it can be
1:43:09automated. Um for example in the B2B
1:43:13sector the the the job that is uh
1:43:18requested with the lot of uh with a big
1:43:22salary sometimes it's forward deployed
1:43:24engineer what is the forward deployed
1:43:26engineer it's someone from a B2B company
1:43:29from u even from frontier models company
1:43:32going into a business and helping them
1:43:35uh get the adoption of the the the
1:43:38agents of the process of the uh
1:43:41automation of the of the uh frontier
1:43:45model. So you need to be like them. You
1:43:47need to be have this forward deployed
1:43:50engineer mindset and be close to clients
1:43:53to help them adopt the the agent, get
1:43:56the feedback and get the feedback loop.
1:43:59Then you can go from prototype to
1:44:01building building to maintaining and
1:44:02then scaling.
1:44:11Okay, how to if you get the product idea
1:44:13test the AI on top internally in case of
1:44:16an almium product if we can validate
1:44:17customer we like the idea then we can
1:44:20add the agent of top streamline the
1:44:22process make the fluid yes that's also
1:44:24that's also a possibility because like
1:44:27agents are accelerators they are not
1:44:30always the response for for the problem
1:44:33so I need I can solve the problem
1:44:35without AI it's even better then I add
1:44:38AI to accelerate and uh we need some
1:44:41real KPIs there. For example, so the
1:44:45what what is the KPIs of agents? We have
1:44:47multiple uh but the more important are u
1:44:52time for example we had the tasks that
1:44:55take uh two weeks or 3 weeks or a month
1:44:59and uh we can help this business doing
1:45:02it in an hour. This is a great KPI and
1:45:05if you can measure it uh this will be
1:45:08really great for your uh business. Uh
1:45:12the second one is cost. Uh for example,
1:45:15we we saw in the market a lot of uh
1:45:17business saying uh if you automate that
1:45:20for for us just build whatever you want
1:45:22because it's really something that cost
1:45:24us a lot of money and uh it's done
1:45:26manually and with a lot of mistakes. Uh
1:45:29so if you can automate this this is
1:45:31really urgent for us. Sometimes there is
1:45:34a cost of doing nothing. Uh we saw in
1:45:37some companies for example they uh they
1:45:40pay money when they don't respond on
1:45:43time for some legal requests. Uh so this
1:45:47is the cost of doing nothing. So if they
1:45:49automate that and they have an agent
1:45:51automating that it means they are saving
1:45:54money uh for for for for this case. So
1:45:57this is the most important KPIs you need
1:45:59to to to get when you are doing of
1:46:02course business but when you are doing
1:46:04personal and when you are doing social
1:46:06one it's about service it's about
1:46:09quality of service when you are doing a
1:46:12society uh agent uh you need to measure
1:46:16the quality of service uh the how it was
1:46:19and how it can be with the AI agent
1:46:30Okay,
1:46:34>> perfect.
1:46:38>> Great. Um,
1:46:42so what we saw this morning, uh, so we
1:46:46had multiple agents how it can help the
1:46:50problem. Morning. Wait Moad, we saw
1:46:53earlier earlier. There was no session
1:46:56this morning.
1:46:58>> Sorry. Sorry. That's um my agent
1:47:01>> participant will say we missed the
1:47:03session. We didn't miss any session.
1:47:05This is what we saw earlier today.
1:47:07>> What we saw earlier? Thank you very much
1:47:09Sara. What we saw earlier today. So
1:47:12about the agent u the action it performs
1:47:16the information and tools. This is
1:47:18really interesting concepts that uh uh
1:47:21that we saw in uh this session. So um
1:47:26many of you um more than 100 not many of
1:47:29you has uh responded to the form um
1:47:32yesterday. So the you filled uh your uh
1:47:37agent instruction uh into the form and
1:47:40uh today what we are waiting from you
1:47:43the takeaway the main takeaway uh have
1:47:45to be the instruction really leveled up
1:47:49uh with what we saw as uh tips for the
1:47:52goal for the context for the tooling and
1:47:54permissions for the output format uh
1:47:57approval and checks and u
1:48:01[clears throat] this is uh we I hope
1:48:03that it changed your your ideas to go
1:48:06from V1 to V2 uh in how to do the
1:48:10instructions. So I'm asking you for
1:48:12those who already filled the form, you
1:48:14can fill the V1 that you did and also uh
1:48:18fill the V2. For those who didn't fill
1:48:21the form, you can go directly to V2 or
1:48:25you can um you can fill V1 and say how
1:48:29it was for you the instruction how it
1:48:32looked like and how today uh session
1:48:35maybe help you enhance it. So you can
1:48:38scan this form and uh fill it please.
1:50:36Well, we have already 245
1:50:39who are responding and the number are
1:50:42increasing. thing.
1:50:59Great.
1:51:32Did I
1:51:44transed?
1:52:07Okay, while we are waiting for the
1:52:08answer to be logged in, uh I'm conscious
1:52:11of the time. So just keep in mind uh
1:52:14that for tomorrow the call is at 3:00
1:52:16and you can find the link on the
1:52:18houseace platform
1:52:23for the checking call. All the question
1:52:25that you uh ask us about the competition
1:52:28will be answered tomorrow. We didn't
1:52:30answer to all of them.
1:52:33So just join us back tomorrow.
1:52:37Um
1:52:40I'll give the floor back to you Mald.
1:52:43>> Yes. Yes. Okay. Let's go next.
1:52:49Okay. So um already having more context
1:52:52about uh your agents. I see that more
1:52:55and more people are feeling.
1:52:58So this was the the sentent before the
1:53:00session. Can we go next please?
1:53:08Okay.
1:53:18So now you have access to fill the V2,
1:53:20right?
1:53:23Okay.
1:53:40So the exercise is um just add your name
1:53:44and your uh instruction as well not just
1:53:47the name.
1:53:48>> So we can select one of them.
1:53:54>> For example, we have
1:53:56okay we have many questions. So use the
1:53:59mentimeter for the version two of your
1:54:05prompt.
1:54:12Okay. Can we see maybe uh some
1:54:20Okay. So, we can see uh lot of you give
1:54:23goals. Uh but of course we need um one
1:54:28of the takeaways of this session is
1:54:31about agents are about tool calling uh
1:54:35and actions. So we need to see more of
1:54:38tool calling in actions. When you think
1:54:40about your agents uh when you think
1:54:44about local agent that will be
1:54:47performing some task inside uh of a
1:54:50company uh what kind of actions it will
1:54:53perform. So uh we wait more of that. I
1:54:57will keep I will keep uh this form open.
1:55:01So in for for the sake of time SA will
1:55:04keep it open and maybe uh you can work
1:55:08at it at ease but we you need to think
1:55:11about uh tools permissions and actions
1:55:15uh that you will be given to your
1:55:18agents.
1:55:20Okay.
1:55:24So I'll be sharing my screen.
1:55:31Um
1:55:34okay.
1:55:37So um we are speaking of course about uh
1:55:40when when we are speaking about agents
1:55:43it's about the tool you are going to
1:55:44give to agents but also we we are also
1:55:48speaking about the tool you are going to
1:55:50build with. uh if we want to get back to
1:55:53the perspective of uh the the challenge.
1:55:56Uh so today's demo we saw uh workflows
1:56:01and uh we we uh we saw four examples of
1:56:06uh workflows. We will see broader
1:56:09autonomy for long longer and delegated
1:56:12tasks. So an agent that can delegate
1:56:15task it about uh broader autonomy and um
1:56:20the session of today was about
1:56:22discovering so the how the agent work
1:56:25how your idea can fit with the with the
1:56:29category uh how it connects with the the
1:56:33second session and third session. So the
1:56:35second session will be about finding
1:56:37opportunities. So we spoke about
1:56:39multiple opportunities we have today. We
1:56:41spoke about local, we spoke about uh uh
1:56:44because I saw many of you choose
1:56:46business but uh it's also an opportunity
1:56:49to to go for other uh other areas uh
1:56:54about personal agents about social
1:56:57agents. There is a lot of things to do
1:57:00and uh the the third one the third
1:57:03workshop will be uh starting in
1:57:05prototyping your first agent and uh
1:57:09moving. So we saw a lot of people having
1:57:11only idea. So the ones having idea
1:57:14moving to prototype and the ones having
1:57:17a prototype moving into a live. Um
1:57:22so um last one last uh form we want you
1:57:27to fill. If you can scan this one
1:57:37and I will share my screen.
1:57:55So the level of delegation to our agents
1:57:58for your projects
1:58:01I'll be interested
1:58:03interested to know That's
1:58:19high confidence to give your agents
1:58:27the right level of delegation.
1:58:33Okay, let's go. Next one.
1:58:56So we spoke about bounded workflows. So
1:59:00what makes our workflow bounded?
1:59:05Is it just calling a personal assistant
1:59:08having some fixed tools and limits
1:59:13or just assume that the model will
1:59:16agree?
1:59:24Great. Great.
1:59:27So we can have a workflow bounded by
1:59:29fixing tools and limits and the review
1:59:32gates.
1:59:41Awesome. Awesome. Congrats team. So this
1:59:44is the leaderboard from all the quiz you
1:59:47filled in this session. And uh team,
1:59:51congrats.
1:59:52Uh if you can write also in the chat,
1:59:55congrats for being uh the one who
1:59:57responded the faster and also with the
1:59:59right answers.
2:00:03[laughter]
2:00:05Great. Congrats.
2:00:09That's it for me. Thank you very much. I
2:00:12give speaking to Sah.
2:00:15>> Thank you very much, Morad. And thank
2:00:17you for everyone uh for following and
2:00:20for asking question. We have received an
2:00:24enormous uh number of questions. Some of
2:00:28them we've answered, some of them
2:00:30related to the competition. We will
2:00:32answer tomorrow
2:00:34during our uh call. So feel free to
2:00:37visit the page on housepace on the
2:00:40community chat and on the workshop uh
2:00:43page as well for you to have the link
2:00:46for the call tomorrow under live under
2:00:49live workshop. You will also have access
2:00:52to the two other workshops we are
2:00:54running during this um agentic AI
2:00:57discovery session. So these are three
2:01:00intro introductory uh session on agentic
2:01:03AI. The next one is next week at the
2:01:06same time from 3 to 5:00 pm GST. So we
2:01:11look forward to seeing you tomorrow
2:01:13during the catchup call and on next week
2:01:15as well for you to learn more. The
2:01:18session will be recorded and available
2:01:22within 48 hours on the Housepace
2:01:24platform. time for our team to edit and
2:01:28to uh bring it to to you uh on the house
2:01:31platform as well as the exercise sheet
2:01:35that will be provided at the same time.
2:01:38For now, I wish you a lovely evening.
2:01:40Thank you everyone for being here and I
2:01:43look forward to speaking to you
2:01:44tomorrow. Thank you Morad for your great
2:01:46session and thank you everyone who
2:01:49participated to the session. Goodbye.

About this transcript

This page contains the full transcript of Discover Agentic AI | Workshop 1 | Create AI Agents Championship by Create AI Agents, generated from the public captions YouTube serves with the video. The transcript has 15,491 words across 2,192 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.