YT YouTube2Text

YouTube2Text → Transcripts → 8 сентября 2026 г.

8 сентября 2026 г. — Transcript

by Andrei · 9,073 words · 1,283 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00section. So, feel free to reach out.
0:02>> Hey everyone, this is Sean. Today, I
0:03want to talk about how to build a strong
0:05AI agent system. So, recently I've been
0:07spending a lot of time talking to
0:08customers, investors, and other AI
0:10founders. And I've got a lot of quite
0:12good feedback on how to build AI agent
0:14systems. And uh not only from how people
0:17use products, but also from how people
0:20build products. Um and I think that I
0:22have collected eight of my top um
0:26experiences that I have collected both
0:27from my own experience and from what I
0:29have discussed with others. So as you
0:31can see I've already listed on my
0:32screen. I'm going to walk you through
0:34them one by one. Uh and hopefully I'll
0:36be helpful for your own uh product
0:38building. And if you're a founder
0:39yourself, I think it'll be very good
0:41reminder because um this is basically
0:43what you're going to need to think about
0:45every day when you iterate your product
0:47and launch the next features. Let's jump
0:49right into it. The oneline summary that
0:51I prepared for myself is that if you
0:54want to build really strong AI agent
0:56systems, you basically need to build
0:57agents that are example rich,
1:00self-improving through human feedback
1:02and simple for customers and verbose in
1:05natural language prompting. I'll explain
1:07them why and I'll go through them uh in
1:10my list down below. Okay,
1:13so the first one I think that's super
1:16important is that smart defaults are
1:19more important than infinite
1:20customization.
1:22What I mean by that is that often times
1:24the customers are usually a lot worse
1:27prompt engineers than you because if
1:30imagine you're a user right and then you
1:31jump into chacht and then chach asks you
1:34to write some prompts for how they
1:36should structure their uh memories for
1:38example which is one of the top features
1:40why people stay on chacht you probably
1:43be thinking about okay what should I be
1:45writing about uh should I say okay keep
1:48everything that I said about my school
1:50about my work as you know some kind of
1:53important information don't remember
1:56anything I talk about you know in terms
1:58of my family about my privacy you need
2:01to think about these like very deeply
2:02otherwise it could be making some
2:04mistakes right when one like similarity
2:07here is that if you walk into Walmart
2:09for example there are like 20 or 30
2:12different brands of shampoo you have no
2:14idea which one to choose from right and
2:16choosing is already easier than typing
2:18your own prompt so often times your
2:20customers might not even know what
2:22exactly they should be writing if you
2:23give them too much control on your
2:25prompt, right? And then what you really
2:27should be doing is that you want to
2:29build a strong defaults that you already
2:31know that works in your domain. Say for
2:34instance, if you're building a law LLM
2:36like Harvey and you're probably a lawyer
2:38yourself and you know exactly what you
2:40should be writing, right? Let's say
2:42you're drafting a new, you know, like
2:44petition for someone who is going to
2:47apply for a green card in the US. Then
2:50the problems that you prepare will
2:51probably be involving like the past 10
2:53years of experience that you have you
2:55know practiced law in immigration law
2:57right and if you let someone who just
2:59you know started working in H-1B visa to
3:03draft their own agents it's probably
3:06less relevant okay so you need to be
3:09having a strong opinion on what exactly
3:12um works in your own domain and then you
3:15also want to keep some blackbox
3:17stability over there because that's kind
3:20of a company secret but also you kind of
3:22want don't want people to mess up with
3:23it because that's based on what you have
3:25learned in the past right and eventually
3:27you want to optimize for the outcome of
3:29the user rather than something that is
3:31infinitely customizable this is actually
3:34a feedback that I got from one of our
3:35investors and he basically suggested
3:37that hey Sean when you're building these
3:39like AI agents make sure that uh you
3:42brought in a lot of your own experience
3:44which in my case is that we I used to
3:47work at Google business messaging where
3:49we enabled more than 10 million
3:50businesses messaging directly with their
3:52customers. So now we're building
3:53something similar. We're building
3:54something that allows all these
3:56businesses to chat with an agent uh
3:59sorry allow their customers to chat with
4:00an agent eventually turn these uh
4:02conversations into a structured leads.
4:05So we are going to embed a lot of what
4:06we have already known about this
4:08industry into our agent. So that is why
4:10I got this feedback which I think is
4:12super valuable. Okay, move on. Next one.
4:15Number two, meet users where they are.
4:18Don't change how users work today.
4:21Basically, it means do not force people
4:23to change their behaviors. Uh imagine, I
4:25don't know, like if you are a
4:26left-handed person and then I ask you to
4:29play tennis with your right hand. Maybe
4:30that's not a good example, but just
4:32imagine you're like terrible at playing
4:33tennis with your right hand, right? And
4:35then I ask you to do that. It just
4:37doesn't make sense. The better way to to
4:39win a game is basically as a coach, I
4:42would try to, you know, assist you with
4:43your left hand uh playing tennis rather
4:46than telling you, hey, right hand is
4:48much better because everybody's doing
4:49this. We have this new tool. Don't do
4:51that. Okay? And especially if you're
4:53building a B2B SAS, what you're not
4:55going to be able to avoid is that most
4:57of your customers probably using uh
4:59Google sheet or Excel or spreadsheets,
5:02right? And a lot of the startups, if
5:03you're targeting startups, are using
5:04notion. So what you really need to think
5:06about is that how can we build our Asian
5:09flow around these existing tools, right?
5:11Follow those familiar patterns rather
5:13than forcing a complicated new view for
5:15the users. And I think that's incredibly
5:18important to keep in mind all the time
5:20because as founders or as product
5:22builders ourselves, it's very easy for
5:25us to sort of fall into the product that
5:27I built because I feel like, oh, this is
5:30one of the this beautiful thing that I
5:32created. But in reality, your user might
5:34just feel like, "Okay, I I I I see a new
5:37dashboard. I see a bunch of new tunch of
5:40new like features. They look very smart,
5:42but I my entire company's data live on
5:45Google sheet." And if you don't adjust
5:47for the workflows that your customers
5:49are already using like spreadsheets,
5:51what you're going to end up with is
5:52basically their budget will be spent on
5:54your product for maybe a year or a
5:56quarter or something and your product
5:58just left there. nobody is touching it
5:59in the company and eventually they will
6:01stop renewing, right? So that's not the
6:03situation that we want to be in if we're
6:05a B2B founder. So yeah, that's that's
6:08why I think this is super important. And
6:11the third one is basically vertical
6:13first strategy. Okay, so this one is
6:15also advice that I received from one of
6:17the investors I spoke with. They
6:20basically are saying that the most
6:22effective way for a founder to start
6:25expanding a product is that you want to
6:28do you want to build these agents around
6:30a very specific vertical and really
6:32really double down in it. Right? And
6:34then what they're saying is that a very
6:36high volume detailed domain specific
6:39examples will make the model stop doing
6:41stupid stuff. Okay? So this is actually
6:44another AI founder friend that I spoke
6:46with also told me about this that he
6:48would rather put like 20 or 30 examples
6:50of exactly how an agent should be
6:53interacting with the user if it's a
6:55chatbot or how exactly an agent should
6:57be dealing with some like like um JSON
7:00files or or PDFs or stuff like that so
7:03that it will actually know you know like
7:07putting these 20 30 examples in so that
7:09it will actually know how to behave and
7:12perform in specific situations. And
7:14because that you're very specific,
7:16you're very in a very specific vertical,
7:18what happens is that your agents will be
7:21feeling like it's doing one thing at a
7:22time, right? One thing that's really
7:24really like spec that's really really
7:26special about LM that it's like a
7:28generic expert. It knows everything. And
7:31especially these days, a lot of LLMs are
7:33using this mixture of expert kind of
7:36model kind of architecture for their
7:37LLMs. Meaning that they're like they
7:39know many different domains since
7:42therefore they're like experts in many
7:44different aspects of life. And you need
7:46to tell it that hey you yours job is
7:49simply to be the expert in this one
7:50thing and this very spec expert of this
7:54vertical and then you'll perform so much
7:57better than like say doing it in a
8:00horizontal level like from the
8:02beginning. Okay. And then later if you
8:04realize that there's a part of market
8:05fit in that domain and also your agent
8:07is doing a really good job you have two
8:08ways. you either like double down on it
8:10and if it's a big enough market then
8:12it's already a venture scalable business
8:14and if it's not then you can think about
8:16expanding horizontally right you can
8:18follow the same architecture of mixture
8:20of experts meaning you could build a
8:23mixture of vertical agents and let them
8:26deal with very different things right so
8:29I think that's a very helpful advice to
8:33keep in mind because uh as builders or
8:36as product founders it's very easy or
8:38common for to think that hey LM can't
8:40answer any questions. Let me just use
8:41one agent for everything. I don't think
8:43that's the best practice in reality.
8:45Okay. Number four is that examples need
8:49to be carefully balanced when you feed
8:51examples to the AI agents. Basically, if
8:54you use too few examples, then the model
8:57would think that there are only a few
8:58ways to solve the problem. You might
9:00actually restrict them. And if you have
9:02too many of the wrong kind of examples,
9:05then the model might overfit to some bad
9:08patterns. And this idea is very similar
9:10to how we were approaching machine
9:11learning about the data sample size and
9:13about the parameter size, right? But in
9:15the LN world today, it's more about how
9:18do you like structure these things in
9:20natural language and then provide them
9:22with a good balance of very solid
9:24examples so that your agents for this
9:26very vertical will be able to understand
9:28how to solve that task. Number five is
9:31that you need to write the workflows as
9:33explicit natural language specs. Okay.
9:37So basically what I have gathered
9:40throughout the information online is
9:42that multi-line stepbystep natural
9:45language instructions will minimize
9:47hallucination and give an agent stable
9:49predictable behavior. Okay. What we mean
9:51by that is that uh sometimes we might
9:53just want to throw a table to LMS and
9:56say okay read this or do this and that
9:58right or might maybe you know we might
10:00throw a bunch of files to it we might
10:02just give it if you're more technical
10:04give it a bunch of JSON
10:06what turns out to be more effective is
10:08that if you can explain in natural
10:10language okay here's the first step you
10:13should be doing ABCD and then if you see
10:16this do this if you see that do this and
10:19then the next step do this and under
10:21what situation you should be using this
10:23tool that API this tool right and here's
10:26the explanation of these toolboxes like
10:28use the stripe API for payments use the
10:31superbase API for fetching data um under
10:35the situation where you know if this
10:37person has authorized yada yada yada you
10:40want to treat them as like a like an
10:43intern your company or as you know like
10:46if you're a product manager as if you're
10:48writing some product documentations
10:50You don't want to just throw a dock at
10:53people. You want to explain to them
10:55exactly what they should be doing,
10:57right? So somehow I think LLM's pick
10:59that up from human beings and then and
11:01also because it's a it's a language
11:03model, right? So explaining in words um
11:06and explaining them in very structured
11:08concise manner with logic and with
11:12experience really helps with the
11:14performance of AI agent system. Number
11:16six, add human the loop correction
11:18loops. Have some kind of approval
11:21rejection cycles so that it acts like a
11:23micro training rounds. All right, so if
11:26you have some good correction and then
11:28you have good improvement and then your
11:30stability endures for longer. Okay,
11:33I think this one specifically applies to
11:36customer support which is these days
11:38people call it CX agent, customer
11:40experience agent.
11:42There are a lot of examples out there.
11:43There are Finn, there are ADA, there's
11:47another company called um uh Sierra AI.
11:51They provide a very uh strong customer
11:55support, customer experience Asian
11:57services out there. What they normally
12:00would always have is that there's a way
12:02to trigger humans to be looped in. Say
12:05for example, Door Dash. If Doash is
12:07using some of these like AI agent
12:08services, there's always like a team
12:11being maintained right there who are
12:13going to type in like answers to to
12:15customers. And when we build these AI
12:18agent system, we always want to automate
12:20the whole thing. But that's not the
12:21reality. In reality, customers would be
12:24like, you know, they would just directly
12:26be requesting for a human to be looked
12:28in, right? Or a customer might be, you
12:31know, super angry and then agents are
12:33still talking like a robot. In these
12:35kind of situations, what you should
12:36should do to maintain a good uh customer
12:39satisfaction score or you know just for
12:42your own retention sake, you should uh
12:45looping humans at that point and then
12:48you should have some kind of mechanism
12:50to trigger that and also you should
12:53train on these good and bad examples so
12:56that your agents in the future will
12:58understand how would I address a
13:00situation where customers are super
13:01angry or when they're requesting for a
13:04human or how do we you know reduce these
13:05situations. These are super important
13:07lessons to be learned and it's very
13:10important to be mindful of that when you
13:11write prompts. Number seven, give your
13:14agent a single memory layer. I think
13:17this is more or less related to context
13:19engineering across a multi- aent system.
13:22What this means is that so for example,
13:24if you're building a a SAS where there
13:26are a bunch of agents, for example, I
13:28think now we're using notion, right? In
13:30notion there there are agents. I'm
13:31pretty sure what they're doing is that
13:32they have a multi- aent system. Maybe
13:34there's some agent are just for, you
13:37know, polishing the paragraph you just
13:38wrote. There's some other agents who are
13:40like reading the entire doc and then
13:42give you a summary. And then if you're
13:44like on the same page and you're talking
13:46to one agent or you're doing like
13:48multiple things at the same time,
13:51they should I'm not sure, Ivan, tell me
13:54if you're watching this. They should
13:58keep one context layer somewhere so that
14:01every agent will have access to the
14:05historical usage of agent in the entire
14:08system either it being chat history
14:11being how like Sean highlighted
14:13something and then asked a question
14:15there what's the context there these
14:18things will be um considered as a single
14:21memory layer or you know you have data
14:23tables there you have calendar events
14:25Um, if you're like making a tool that's
14:27helping people to schedule meetings or
14:29having a voice agent that will, you
14:31know, schedule meetings, book
14:32appointments for for dental offices,
14:34stuff like that, having a single memory
14:36layer is always super important because
14:38that just simply solves the problem of,
14:41you know, some parts of your product
14:43just don't remember what the user did uh
14:45previously, right? And if you're say for
14:47example if you're like uh building an
14:48email agent then this email agent should
14:51have access to some sort of you know
14:52calendar events that uh you previously
14:54scheduled for them if you have another
14:56voice agent right there. Okay. So it
14:59might sound a little vague if you
15:00haven't built these things before but if
15:02you did then you will know what I'm
15:04talking about. Um, if it's the first
15:06time for you to heard about what is a
15:08memory layer, uh, try to check out, you
15:10know, things like how do you build up
15:11some context, how do you build up some
15:14shared states for agents? Search these
15:16things up and then I'm sure AI will give
15:18you a really good answer as well. Last
15:21but not least, which is something that
15:24YC partners would emphasize all the
15:26time, which is eval. It's important to
15:29add a continuous evals to measure and
15:31enforce reliability. Okay. So what in
15:34reality agents regress very easily and
15:37automated evals across workflow
15:39scenarios and edge cases ensure
15:40improvements are real. Let me give an
15:43example. If you say are building an
15:45email agent and this email agent is
15:46supposed to check the headline, check
15:48the subject, check uh the content, make
15:51sure you have the right signature, make
15:54sure that it read the previous context
15:56and mention anything as a as a greeting
15:59like as as as context that you know how
16:01to greet this person. if it's a
16:03follow-up and then what happens is that
16:05sometimes agents might skip the steps
16:08depends on you know today maybe claude
16:10is not performing well tomorrow chbd is
16:13like replacing their old model with some
16:15updated ones anything could happen right
16:17so what you should do is that you kind
16:19of have to have the system there to do
16:21the evaluation and be like okay before
16:23you hit send of any new emails make sure
16:26you check that each one of these steps
16:28are fixed or are like fulfilled and you
16:31have some kind of like code there to
16:33make sure that it's actually, you know,
16:35uh filling in the the information
16:38requested right before you send the
16:40email. That's kind of the eval system, a
16:41very simple one that I could come up
16:43with right now. And uh having these
16:46evals that are very specific to your own
16:50agentic system or your own agentic
16:53workflow
16:54is
16:56kind of one of the most important molds
16:58out there for AI startups uh or AI
17:00products because you know your product,
17:04you know how to make it better. You also
17:06know that this is not going wrong
17:08because you have this um very very
17:11strong like like um like security system
17:14or stability maint maintenance system
17:17out there so that your agents do not
17:19regress. Okay. And I forgot to mention
17:21another very important one is a single
17:23memory layer which is you know like
17:25that's honestly one of my uh founder
17:28friends recently just said that he's not
17:29switching to Gemini even though Gemini
17:31is really good and the reason is because
17:34Tachib has got his entire past two years
17:37of history and uh he's more comfortable
17:40just talking with Chachi because he
17:42doesn't need to explain again what his
17:44company really is what was that last
17:46contract they just signed stuff like
17:48that okay these are super important
17:50modes for your company.
17:52Last but not least, lens are super
17:55capable. If an agent messes things up,
17:57it's almost always almost almost always
18:00your prompt's fault instead of the
18:02model's fault. So, this is coming from
18:04one of my founder friends quote, "If
18:06Claude can learn code bases, can learn
18:10how to write code, your agent has no
18:12excuse. You should give them as much of
18:13an example as possible." And the example
18:16was that code uh Claude leaked their um
18:20uh uh prompts out there, right? So
18:24basically they just put a ton of
18:25examples in their coding uh tasks.
18:28That's why they are they know how to
18:30write code very well. Okay? So if cloud
18:32can learn codebase, your agent has no
18:34excuse. Your agent should be better
18:36prompt so that it does exactly what you
18:39want it to do. Okay?
18:41So I know I've been talking for a long
18:44time but let's come back and recap this.
18:46Essentially if you want to build agents
18:49AI agent system that is very strong what
18:51you got to do is build agents that are
18:53example rich self-improving through
18:55human feedback simple for customers and
18:58verbose and natural language prompting.
19:00And I guess maybe here h I should
19:02probably include uh with a memory layer
19:05as well. uh let's just say with a memory
19:09layer and a bounce.
19:11Okay.
19:14Yep. That's basically a very highlevel
19:18summary of how to build strong agent
19:20system based on my personal experience
19:22based on the conversations I had with
19:24customers with AI founders and
19:26investors. Hopefully this is helpful for
19:28you and uh let me know if you have any
19:30questions. I'll see you next time.
19:32Thanks. Hey everyone, this is John. So
19:34today, let's walk through how to build
19:35an a gentic rack system like a pro.
19:38We're going to walk through a quick
19:40system design over here and then we're
19:41going to jump right into a live demo
19:43that will not only have an embedding in
19:46a vector database for retrievalss, but
19:48we will also show you how to use some
19:49tools like setting up Google calendar,
19:52setting up a Gmail, and last but not
19:54least, I'm going to show you a
19:56open-source um GitHub repo for identic
19:59rock that will allow us to set up the
20:01entire codebase with superbase, Python
20:04backend, and Google cloud. So, let's get
20:06started. Firstly, and this is a system
20:08design for an agentic rack. For some of
20:11you who watched my previous video about
20:12a YouTube rack system. This is basically
20:14an extension. We introduced this new
20:16thing called agentic tool. The user will
20:18have a user channel that will either you
20:20know talk through u a web chat, an email
20:23or a WhatsApp. And eventually it's going
20:25to interact with an AI agent that sort
20:27of is doing the communications over
20:29there, right? And then according to our
20:31previous video, we basically allow this
20:33agent to query some data from a vector
20:35database which in our case we're using
20:37superbase. And this vector database
20:39essentially is like splitting some
20:41original documents, turn them into
20:43smaller chunks and then turn those words
20:45into embeddings. Embeddings is basically
20:47highdimensional vector uh that will
20:49represent words in numbers. Eventually
20:52this AI agent will sort of check the
20:54policies, check what kind of products
20:55they have. That's the traditional rag
20:58we're talking about today. We're going
20:59to show you this very quick demo of
21:01talking to an AI agent and actually
21:03letting it schedule a call with us and
21:06eventually sending us an email. And I'm
21:09very excited to show you the demo right
21:10now. This is a local host 8000/hat.
21:14And uh let's just jump right in and see
21:16what happens. And you can see that
21:17currently we have this thing called a
21:20scheduling time zone v1. I sort of set
21:23up the uh time zone for availabilities
21:26for this uh AI agent to respond to a
21:28consumer. So if I say I want to schedule
21:32a call with your manager to talk about a
21:37wholesale deal. Let's see what it says.
21:41All right. Now firstly it asked me to
21:43schedule a call. Before that we should
21:45uh they asked me about the preferred
21:47date and time for the call. And I'm just
21:49going to say, okay, can we do 6 a.m. EST
21:53tomorrow?
21:56Now, ideally, it should check um the
21:59policy over here, and it say that the
22:00earliest available time slot is 7 a.m.,
22:02which is true. 7 a.m. right here. Okay.
22:05So, it cannot allow me to schedule time.
22:08So, let's just say, okay, cool. Let's do
22:1110:00 a.m. EST tomorrow. My email is
22:15Sasha.
22:17atgmail.com
22:20and meanwhile let's just show the
22:22calendar right here. Okay, let's send
22:24this.
22:27Cool. So now it has set up the call for
22:29me. You can see that it popped up real
22:31quick immediately, right? Look at this.
22:33It's already here. Uh sent from Sean at
22:35automatis.io. Automatics is our startup
22:37by the way, which is a sales leads
22:38manager for made to order products. Feel
22:40free to check it out. automatus.io. send
22:42me a summary email as a heads up. So,
22:49you can see that in my mailbox, I
22:51already received this calendar and I can
22:52say yes to it. And if I come back, I
22:55just got another one saying that uh it's
22:58a reminder for upcoming call, which is
23:00literally what we asked for. Okay, so
23:02this agentic tool calling is working.
23:04This is incredible. What happened was
23:06that we asked uh can we do 6 a.m. and
23:08then it checked the policy. They said
23:10that no, 7:00 a.m. is the earliest time.
23:13So, it didn't even do the tool calling.
23:14And then later, I say, "Okay, let's do
23:1610:00 a.m. tomorrow." And immediately
23:17understood and I gave it my email and
23:20then he also, let's see, you also know
23:22that we're doing a wholesale deal
23:24discussion. So, it kind of also has the
23:26context and eventually asked it to send
23:28an email as a heads up. He also sent me
23:29an email as a heads up. So, this is a
23:31full cycle of a aentic rag system that
23:35allows this AI agent to talk to a
23:37consumer. Okay. So now let's uh jump
23:40right into the code and see how it
23:42actually worked. So I prepared this
23:43GitHub called Yt-agentic-rag
23:46under shenan. You guys can all have
23:48access to it. github.com/shencht
23:51agenticrag. I'm going to link the
23:53YouTube link here. If you like the
23:54video, you can click on buy me a coffee
23:56to support us. U what I really wanted to
24:00show you is this current project setup.
24:03We set up this aentic system in a fast
24:05API back end. Just like previously, we
24:08actually had another repo called
24:10YT-Rack. Let me just turn it up for you.
24:13This is YT-Rack. Got 33 stars so far.
24:16What this one did was simply vector the
24:18database for the rack. And there's no
24:20tool calling at all. And this time we're
24:22going to dive into how do we combine
24:23both of them so that the AI agent has
24:25access to not only the knowledge but
24:27also the actions. Okay. And uh you can
24:31see that we have a comparison here that
24:33um previously we also have vector search
24:36and rack uh for Q&A but this time we
24:38we're getting tool calling agentic
24:40reason loop calendar scheduling email
24:42sending multi-turn chat history
24:44multi-step actions. So if you scroll
24:46down a little bit you can see that
24:47there's a step one and then you
24:49basically just get clone this repo and
24:51uh set up some Python environment. All
24:53you got to do is just copy this, turn on
24:55your clock code, cursor, VS code,
24:57whatever. I use cursor. So, let's try
24:59the cursor version. So, all you got to
25:01do is you literally just find somewhere
25:02empty and then paste this whole thing in
25:04and then hit enter. It will start a new
25:06project for you. And then the next step
25:08is that you need to set up your
25:10Superbase. Basically, for anyone who's
25:12not familiar with it, Sub is one of the
25:13most popular agent related database. Uh
25:17and then they also have their own vector
25:19embeddings for uh retrieval augmented
25:21generation tools. So it's very handy,
25:24very helpful. Uh you just got to create
25:26a new project, set up the password, set
25:28up a new region, and um uh all you need
25:31to do is to copy these APIs and the
25:34project URLs. And um if you go back to
25:38the code, you can see that we got this
25:40m.example.
25:42What you need to do is that you need to
25:43copy this, turn it into a m and then
25:46replace these superbase URL anon key and
25:50service control keys with the real ones.
25:52Okay, I'll leave this to you guys and
25:54feel free to check my previous videos
25:56for how to set up superbase. I've got
25:57plenty of examples. So after superbase,
26:00uh there's another thing which is you
26:02need to set up the database in our code.
26:04We have a section called SQL and you can
26:06click on initial superbase SQL and you
26:09can see that we're showing you how to
26:11create the tables. All you got to do is
26:12just command A, command C if you're
26:14using Apple. Come to superbase. Come to
26:17SQL editor. Click on plus. Click on new.
26:20And then just paste it here and run it.
26:22That's it. And then you're going to set
26:23up the tables just like me. All right.
26:25On the sidebar table editor, you can see
26:28that I was showing you the example of
26:29the policies and all of them. We have
26:32some initial setup just for you. If you
26:34run this project after you set up the
26:36superbase, what happens is that you need
26:38to also set up the openi keys and then
26:41eventually set up the Google cloud
26:43because we're using Google calendar and
26:45Gmail. So it's going to be much easier.
26:47So what happens is that you need to set
26:49up the thing this thing called a service
26:51account. What the service account does
26:53is you're basically telling Google that
26:55uh this app this app is allowed to use
26:59my Gmail to send calendar invite to this
27:03extra Gmail externally. Okay. And then
27:06the way you do that is just you need to
27:08come to
27:09console.cloud.google.com/appi/credentials
27:12or you can just navigate to the lefth
27:14hand sidebar API and services and click
27:16on credentials. Okay. And then you'll be
27:19able to um land on this page where you
27:22can click on create credentials plus a
27:25new service account. Okay. And a service
27:28account is basically you can handle
27:29everything in the back end without any
27:31authorizations anymore. It just you're
27:33giving it a secret or access to your
27:34project. Okay. So you just got to put in
27:36your account number uh sorry account
27:38name and then put in a new ID uh and
27:41then put in the description and just
27:43follow the follow the instruction and
27:44finish the whole thing. I have
27:46documented everything here. So you just
27:48need to literally follow this step,
27:49right? Give it as a name, agentic rack
27:51service. Um, and then eventually you
27:54need to download the service key to your
27:57local folder called credentials/service
28:00account.json which is right here. Uh, we
28:03have a folder called credentials. You
28:05can see you need to download that key
28:07over here and call it
28:08service_acount.json.
28:10Okay, that's all you got to do. And if
28:13you're deploying this to Google Cloud,
28:15you also need to set up something else.
28:17All right. But um other than that, you
28:20just need to make sure that you entered
28:22this ooth scope um into into your Google
28:26client ID. All right. And then just
28:28authorize the whole thing. And then you
28:30just need to set up the environment
28:32variable. And that's about it. So in our
28:34codebase, uh we have an app folder in
28:37which you have the agent folder. And
28:40then our main agent is over here called
28:42ultrater. Okay. I'm just going to keep
28:44the file which is all the changes and
28:46then what this one does is that this is
28:48the central agent that will help you to
28:51retrieve relevant contacts from the rack
28:54and build the message system um doing
28:56the functional calling tool calling and
28:58eventually generate the final response
29:01um with you know the output. Okay. And
29:04then this one is is being fed with the
29:06tool folder in which you can see that
29:08we've got for example this uh calendar
29:11tool. Okay. This calendar tool basically
29:14has um an automatic Google meet link
29:17generation multiple attendees time zone
29:19support rag informed duration. For
29:22example, it's scheduled 30 minutes
29:23because in our policy it say 30 minutes.
29:26Okay, so this is very handy. And then in
29:28the service folder, we also have the
29:30rest of the things just like the
29:31previous uh video where we basically
29:34tell you tell the agent how to chat uh
29:36how to chunk the documents into smaller
29:39chunks and how to uh run the embeddings
29:41using a open AI. This demo is literally
29:44focused on calendar tool and email
29:46tools. All right. And for the calendar
29:48tool, essentially it's a it's a new uh
29:52service account that you created using
29:54Google cloud and uh eventually you're
29:56going to execute it um by sharing you
30:00know exactly what's the start time, end
30:01time, description, attendees and time
30:03zones. Okay, feel free to play around
30:05with this repo. It should be pretty
30:06handy for you already. Hope this makes
30:08sense. Let me know if you have any
30:10questions. This is probably one of the
30:11simplest agentic rack system full stack
30:13from system design to code to deployment
30:17on the internet. Let me know if you have
30:18any feedback and I'll see you next time.
30:20Hey everyone, this is Sean. So today
30:21we're launching Automanas AI agent for
30:24businesses who are selling customized
30:25goods ranging from cabinets, solar
30:28panels, EV cars to module based
30:30softwares. So a lot of business we're
30:32speaking with are either qualifying
30:33their customers through WhatsApp, SMS or
30:36emails. Many of these businesses are
30:38talking to 20 to 50 customers at the
30:40same time and maybe like 20% of them
30:42would actually convert into a final
30:43conversation and then turn into actual
30:45leads. So a lot of time are wasted on
30:47qualifying the customer and eventually
30:49manually typing things into CRM. So
30:51we're going to change that. Okay. So
30:53today I'm going to show you how a
30:54consumer would interact with automatis
30:56WhatsApp agent and then eventually we
30:58will configure the agent and then turn
31:00it into a structure leads. Let's get
31:02started. So essentially as you can see
31:03that firstly we ask this question. and I
31:05I say hi hey this is Sean how are you
31:07and they say that this is the automatis
31:09AI receptionist so on the right hand
31:10side you might be wondering what's going
31:12on here this is an agent that we train
31:14for um this specific business and it's
31:16called EV Dubai Ltd okay so essentially
31:20this is a customized page where you can
31:22do some multiple choice questions and
31:23then you can talk about like what
31:25restricted topics there could be what
31:27you need to do is that you need to tell
31:28it okay what kind of products are you
31:30selling right for example this EV brand
31:32is selling BYD CL 2022 and then Tesla
31:35Model Y with their own price. And then
31:37you can just quickly test it here and be
31:38like, do you have BY models
31:43and the simulator will be syncing
31:45exactly what you have uploaded to your
31:47knowledge base as well as the things
31:49that you have updated on the main
31:50behavior guidance training page. All
31:52right. So now you say that BYD sales has
31:54$42,000 USD. Okay. You can also modify
31:57the formality and do you want it to
31:59become or more passionate? Eventually
32:01you just got to click on deploy. And
32:02eventually if you come back to overview,
32:04you can see that there's an agent
32:06currently being deployed under this AI
32:08agent hub. We're using the share number
32:10plus one 6506056956
32:13for automat. If I just ask you questions
32:14and be like, okay, um, do you have a by
32:19model
32:21or an EV? I want to buy one. Right? Now,
32:24automatically doesn't know who you are
32:25and who you're looking for, right? And
32:27then after you show the intent, it's
32:29kind of sort of processing in a search
32:30across the entire business's network
32:32that has been connected with Automatis.
32:34And it found out that oh, there's a one
32:35called EV Dubai Ltd, right? And as a
32:38user, if I click on it, what it does is
32:40that it's going to trigger this agent
32:42for me. If I want to exit anytime, it
32:43just type exit anytime. In our knowledge
32:45base, we have these products here. So,
32:47I'm just going to ask it, can I buy the
32:50UID SEO car? Do you have a model? And
32:55what is the price? So as a business you
32:58might have these kind of questions maybe
32:59a thousand times every week right and
33:01then what you don't want to do is that
33:02you want to you don't want to answer all
33:04the questions again and again and maybe
33:05you want to configure your own way to
33:07answer these questions which is all
33:09doable in this behavioral guidance page
33:11right and now it has told me that okay
33:13the by seal is available which is
33:14$42,000 exactly what we put in in the
33:17knowledge base and what we need to do is
33:19that uh yes my budget is 50k
33:25let's get the BYID seal model for 52K.
33:31I'm just being like dumb and answering
33:33things very thoroughly here. And while
33:35this is being communicated as a
33:37business, you can also come back to
33:39leads on the left hand side. And you can
33:40see that there's a there's a lead
33:41appending in the review. Okay. And if
33:43you tap on it, you can see that Sean
33:45Chen just had a recent journal with
33:47WhatsApp with us. Okay. If I click on
33:49view all, you can see that the entire
33:51conversation is here. All right. So, as
33:53a business, I could approve this as a
33:55lead and then let my sales actually pick
33:57it up or I could just cancel it out,
33:59right? So, today I'm going to approve
34:01it. So, if I click on approve and after
34:03you approved it, it's going to land in
34:05the in progress lead. Okay? So, you can
34:07just always come back and check that
34:08this lead is active. Then what if
34:10somebody has followed up and they say I
34:12also want to buy a
34:15Tesla Model Y and I can double my
34:20budget.
34:23What you will be able to see is that
34:25there's an additional appending review
34:27for this new chat that just happened.
34:29Okay. If I just say approve
34:32again, you can see that the new
34:33conversation has been merged to the
34:35existing deal with Shan Chan. All right.
34:37And if I come back to in progress,
34:40expand it, you can see that there are
34:42two conversations at the same time for
34:44the same customers. And I can also just
34:46click on extract to-dos. And after that,
34:48if you cross it out, you can see that
34:49there are a bunch of to-dos here related
34:51to the conversations that the consumer
34:53just had with the business. And it's
34:54just super convenient for you to handle,
34:57say, multiple conversations at the same
34:58time. Okay. And then you might be
35:00wondering, okay, we're using this shared
35:02WhatsApp number. What if there's another
35:04company? What if my business want to
35:06sell something different? This account
35:07is basically selling cabinets. All
35:09right. And this agent is basically
35:12selling cabinets by a different company.
35:14All right. And if we click on to edit,
35:16you can see that there's a cabinet hero
35:18agent over here. As a consumer, I could
35:20come back here and I say exit. And then
35:23automatics will tell me, got it. You've
35:24been disconnected from EV Dubai Ltd.
35:27What business do you want to speak with
35:28today? Just tell me your company name.
35:30We can just let the consumer to talk to
35:31WhatsApp with your business name. Or as
35:34a consumer myself, I might be just
35:35wondering, okay, I want to upgrade my
35:39kitchen
35:41with a very vague intent. And the agent
35:42that's talking to us right now is the AI
35:45receptionist from Automatis. And he
35:47said, "Okay, the kitchen cup really
35:48transformed your space. Are you looking
35:50to replace cabinets or do you want
35:51specific areas?" It doesn't tell me
35:53exactly what business yet because it
35:55wants to collect more intents from the
35:57consumer. And I can say, "Okay, I want a
35:59new cabinet.
36:02Do you know any business selling it?
36:06Might need to scroll down for this chat
36:08every single time. And now you reply
36:10that. Okay, we have this company called
36:11Queenswood. Right, there we go.
36:12Queenswood's right here. If I click on
36:14Queenswood, it send this message back to
36:16Automanas AI agent. And again, we're
36:19connected to Queenswood agent right
36:21here. Right. So, I just say um do you
36:24have any discounts?
36:28And if I come back to knowledge base, I
36:31have already uploaded
36:33um I have already uploaded a 20% off
36:36discount for this green sling shaker
36:38cabinets. And you can see that
36:41immediately knows the contacts that the
36:4320% off is for this specific product.
36:45And I'll just say I want this. Okay,
36:49click on send. Meanwhile, if I come back
36:52to leads, I click on pending review. You
36:54can see that there's a new conversation
36:56here. uh if I expand it, I also have a
36:59WhatsApp conversation just like before.
37:01Similarly, I could approve it from a
37:04different business, right? Once that
37:06business has approved it, it has nothing
37:07to do with the other business. So, we're
37:09sharing the same number over here that
37:11allows any consumer to if they want to
37:14buy certain things, they just literally
37:15explain what they want to automaticize
37:18WhatsApp agent. Eventually, if the agent
37:19that you deployed is connected to it,
37:21the conversation will land in your leads
37:24in the pending review. Okay? And after
37:26you approve it, if you have clearly
37:28explained what product you want, then
37:31it's also going to capture the product
37:32that the consumer was interested in. So
37:34the way to set it up is actually really
37:36simple. You just come to the landing
37:37page of automatics.io
37:39and then you can either try it by
37:42clicking on the chat with us down at the
37:43button here and start conversation,
37:46open WhatsApp,
37:49and then just send the message. you will
37:51be able to trigger the conversation or
37:54you can come back here and um click on
37:57start free trial. We're providing a
37:5830-day free trial for the starter
38:00package for $20 per month if you click
38:02on it and we're already subscribed. So
38:04after you subscribe um by the way it's
38:06completely free for 30 days. You can
38:07cancel at any time if you want in the
38:09settings. Two last things. If you're
38:11interested in setting up your own AI
38:12agent, you can also click on connect
38:15your own WhatsApp and then connect your
38:16WhatsApp over here. All you got to do is
38:19just put in your access token, phone
38:21number ID, and business account ID, and
38:23you'll be able to set up AI agent for
38:25your home. Let me know if this is
38:26helpful. If you're a business who are
38:28selling customized goods, either
38:30physical ones or softwares at scale,
38:32we're very happy to talk to you. And
38:33just email us at [email protected].
38:36Thanks for your attention. I appreciate
38:38this and uh see you next time. Cheers.
38:40Hey, claw code. Can I create a sales
38:43representative agent for my business?
38:45Yes, sir. Let's go. Every AI product and
38:48agency business needs a way to talk to
38:50your customers. And does not matter if
38:52you're technical or not. You don't want
38:54to build this from scratch every single
38:55time. What you really want is a sales
38:58layer that will be the sales interface
38:59to speak to the world. And that's why
39:01we're introducing Automan's MCP server
39:04and API services for you today because
39:06we want you to be able to create your
39:08first sales representative agent within
39:101 minute. Let me show you how it works.
39:11There's a very simple way for you to set
39:13up MCP. If this is the first time to
39:14hear about it, you just need to go to
39:16your profile and click on it and then
39:17come to settings and developers and you
39:19can see that I have a few MCP servers
39:21right here. Uh notion API is one of them
39:23and Almanus is the other one. You can
39:25just click on edit config.
39:27You see it will pop up this link here
39:29and all you need to do is follow the
39:31instructions on the MCV server
39:32development repo here. You just need to
39:34copy this and then open this cloud
39:37desktop config JSON and then just paste
39:39that in. I have multiple NCP servers
39:41that you can see auto is right here. So,
39:43as you can see, after I ask this
39:44question, it's asking me for three
39:46things. Company name, website URL, and
39:48my email. This is because my claw code
39:50has been set up with the MCP server from
39:53Automanas. So, let's say the company
39:54name is calendarly.com. And also, I'm
39:56going to provide my email for
39:58development. I'm just going to hit
39:59enter. So, now you can see that uh Clo
40:02is going to sort of, you know, using
40:04this MCP server again trying to create
40:06the sales agent for me for this company
40:09and my email. is using our backend
40:11services to you know search this
40:13website. Uh it could be your AI product,
40:16could be your agency website. Um to
40:18understand how do you take in inquiries
40:21from your customers? How what kind of
40:23things or solutions are you providing
40:25for your customers? And the whole goal
40:27of this agent that's creating for me is
40:29to make sure that there's no leakage
40:31from customer inquiries when you're not
40:33in your office or when you're, you know,
40:35out there speaking to other clients. As
40:36I promised, within 1 minute, everything
40:38has been created. You can test it with
40:40two ways. You can either click on this
40:42link right here to chat with the agent
40:44and but that agent currently is on the
40:46system server. You don't own it yet. And
40:48at the same time, you must have received
40:50an email which is
40:53something like this. Uh which would
40:55which is asking you to claim your agent
40:57so that you will be able to uh own it in
41:00your own platform. But let's try on
41:01WhatsApp first. You can see in this
41:03email there's a link called test on
41:05WhatsApp. If I click on that, it will
41:07pop up this WhatsApp link that will
41:08allow us to chat with this ultimat right
41:10here. And just enter. I'd like to talk
41:12to Calendarly.
41:14So, it g me a button to click on to
41:16route to Calendarly. But you can see
41:17this little yellow warning here because
41:19that agent doesn't belong to me, right?
41:21Because I'm not from Calendarly. If
41:23you're an employee or CEO from
41:24Calendarly, you'll be able to claim it
41:26using that email link with your own
41:28email domain so that there's no security
41:30issues. This must be verified by you.
41:33What do you do?
41:37All right. It says, "I'm here to help
41:38you discover Calendarly and simplify
41:40your scheduling. Uh, we help you do a
41:42point booking. How much is your
41:44service?"
41:47Cool. Service is like $10, $16, or 15K a
41:50year. We can also ask Clawo to add more
41:53knowledge to it. So, I'm going to say,
41:54can you add a booking link to its
41:56knowledge base, which is my Calendarly?
41:58Let's go.
42:00So now it's using this MCP tool called
42:03add knowledge. All I need to do is just
42:05say allow and then we'll be able to add
42:08this knowledge base to this agent and we
42:10can see if it's working. All right. So
42:12now it's done. So I'm going to say can I
42:15schedule a calendarly booking with you?
42:16Do you have a link?
42:19Cool. It says yes you can schedule
42:20meeting directly using calendarly link.
42:23And if I click on it I'll be able to
42:24book a call with you. All right. So this
42:26agent is completely created based on
42:28claw code MCP and I think everybody can
42:31do it. So now let me show you how to
42:33actually claim this and then you can own
42:34it and then can modify all the knowledge
42:36that this agent should know. Remember we
42:38received this email. Uh there's a link
42:40called claim your agent. Let's just
42:42click on that. All right. And now it's
42:44then going to direct us to this invite
42:46link that will allow you to claim this
42:48agent if you use the same email that you
42:50use to sign up. Cool. You can see that
42:52we have this new agent called Calendarly
42:54Assistant. And if I just click on
42:57deploy.
43:02Okay. Uh because now we're on a free
43:04plan, so you can only deploy one agent
43:05at a time. If I click on edited, so this
43:08company is called Calendarly Assistant,
43:11and we can see that there's goals and
43:14actions. It scraped the most important
43:17information here. And most importantly
43:19there is a knowledge base that has you
43:21know all the information that we have
43:22scraped from this product. Remember we
43:24say the agent was not verified. Now in
43:26the training page you can click on the
43:28sandbox mode and then put in your
43:30website URL to confirm if your emails
43:32match with it. If it does then it's
43:34going to verify it. If it does not then
43:36your agent will remain unverified for
43:38security reasons. Now you might be
43:39wondering where are the conversations.
43:42You just need to move to inbox and click
43:43on it. You can see the entire
43:46conversations that we had between the
43:48clients and the agent. And what you can
43:50do is you can create a lead from this
43:52conversation by clicking on the button.
43:54So what it's doing is that it's
43:55processing this entire conversation
43:57identifying if they're looking for any
43:59products or trying to book a call with
44:00someone. And then eventually you will be
44:02able to um turn it into a structured
44:05lead. Imagine you could do this in
44:06batches, right? If you have like 700 800
44:09inbounds of inquiries over a month, then
44:12um you might be leaking some uh
44:14potential customers by not replying to
44:16them promptly. With this product, you'll
44:18just be able to handle it without much
44:20of a concern. Okay, so now it's almost
44:22done.
44:24Okay, you can click on create view lead.
44:27You can see uh you know lead hub for the
44:30entire conversation in the to-do. You
44:32can see that the customer was trying to
44:33book a call with us. You cross it out.
44:35You can see um this customer is
44:38potentially a enterprise customization
44:40client. So you capture that the value is
44:42about 15K. So this was a quick demo of
44:44how you would use clock code MCP to
44:46build an agent for your business or for
44:48your product immediately. If you are a
44:50developer, you might not only want to
44:52use MCP, but you want to perhaps you
44:54know set it up using APIs. Maybe you're
44:56building a SAS product or an agency
44:57product that require a lot of back and
44:59forth conversations. And while you're
45:00building the website or updating your
45:02product in cursor, in lovable, in claw
45:06code, uh you might want to, you know,
45:07just have access to the API. All right,
45:09we also have that ready for you. You
45:10just need to come to settings and API
45:12access. And you will be able to uh
45:14create a new key here. Let's just say uh
45:18test key that will expire in 30 days.
45:22Create the key. Copy the key. Click
45:25done. And then all you need to do is
45:27just, you know, use this curl. Let's
45:30just try with a real terminal. I'm just
45:31going to copy this in and create this
45:34company. And now this API is going to
45:36create this agent for us.
45:40It says it's successful. And uh in this
45:44case, it should just be directly in your
45:47agent hub
45:49because you're using the API keys. Okay,
45:53there you go. You have this entire agent
45:56here using the API. You don't even have
45:57to claim it. You just need to sign in
45:59first and then create the API key to use
46:01the API key to call the API. We're also
46:03able to allow you to create a web chat.
46:05So, if you come to AI Asian web chat,
46:07you can say let's say Calendarly and
46:10then we can say create a widget. Uh
46:12choose a different color,
46:16right? Put at the bottom right of your
46:18screen. And then you can just test it
46:21right here. It's like hey what is your
46:25service cost? Okay same thing um you can
46:29also click on a WhatsApp to chat with a
46:31WhatsApp agent. So you know this is the
46:33entire process of how you would use MCP
46:36server or APIs to create a sales
46:37representative agent for your business
46:39for a product uh without much pain. And
46:42if you're curious about this I would
46:43really appreciate it if you can come to
46:44our GitHub server and check out this
46:46documentation. If you really like this
46:48product or you just want to support me,
46:50uh give us a star or fork our our repo
46:53and uh download our MCP server on tier.
46:56Uh everybody will get 100 free credits,
46:58meaning 100 free AI generated responses
47:00for your clients and uh I would love to
47:02hear any feedback and happy to chat
47:04more. Oh, also like this video,
47:06subscribe if you can and uh we have our
47:09Discord server down below in the link.
47:12Uh happy to discuss more in our channel
47:14on Discord. Thank you so

About this transcript

This page contains the full transcript of 8 сентября 2026 г. by Andrei, generated from the public captions YouTube serves with the video. The transcript has 9,073 words across 1,283 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.