YT YouTube2Text

YouTube2Text → Transcripts → Crea agentes de IA en Claude que trabajan por ti en 60 minutos.

Crea agentes de IA en Claude que trabajan por ti en 60 minutos. — Transcript

by Aztec Lab · 11,448 words · 1,646 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:03The chart you are seeing on screen is
0:05the number of Google searches for the
0:07term "AI agent" over the last 5 years.
0:13Notice that in 2022, there was
0:14practically no talk about this topic.
0:172023, 2024, and 2025 as well. And
0:21starting in mid-2025, this began to
0:24explode. Look, we are going through the
0:28period with the most interest in
0:30artificial intelligence agents. And
0:35it's not for nothing, because AI agents
0:37are some of the most powerful things I
0:39have ever experienced in my life and
0:41have the greatest potential to truly
0:43transform the economy and our personal
0:45lives. But there is a problem, which is
0:52that no one has explained what the hell
0:54an AI agent is, how they are created,
0:56how they are built, how they are used,
0:58and most importantly, how can I make
1:00them actually help me be more efficient
1:01, work less on boring tasks, and
1:03generally automate my work? In this
1:08video, I am going to explain to you
1:10what an AI agent really is, no beating
1:12around the bush, no machine drawings,
1:14none of that. What is an artificial
1:17intelligence agent? How can you, with a
1:22single prompt—a text I will give you
1:24—start creating, training, and
1:26configuring your own AI agents, and how
1:28can you use them to truly generate
1:30value and make your lives and jobs much
1:32easier? So, stay until the end because
1:37not only am I going to show you how
1:39this works, but I will also give you
1:40all the resources you need so you can
1:42do it yourselves. This is going to be a
1:45complete class on AI agents and agentic
1:47workflows, and I swear that after this
1:49class, you will have everything—
1:51absolutely everything—to start
1:53training your own agents today. Oh, if
1:55you don't know who I am, my name is
1:57Martin Vazquez. I am the co-founder of
2:00a company called ASTEC, through which
2:02we have helped tens of thousands of
2:04Latin Americans learn about artificial
2:06intelligence, vibe coding, building web
2:08applications, and using AI agents. And
2:12we have worked with some of the most
2:14important companies in the region,
2:16helping them create AI-based solutions,
2:18automate processes, and generally be
2:19much more efficient, both they and
2:21their teams, using artificial
2:22intelligence and technology. So, if you
2:28also want to take advantage of
2:29everything happening in the field of
2:31artificial intelligence, don't want to
2:33get left behind, don't want to miss
2:34this revolution, and want to learn from
2:36a team of people who explain things for
2:38regular folks without unnecessary
2:40technical jargon, subscribe to our
2:42channel, hit the bell icon so you're
2:43always notified when we have a new
2:45video, and if you like this video or
2:47have questions, leave a comment. Oh,
2:50and of course, don't forget to hit like
2:52. It really helps us so that more
2:54people learn about artificial
2:55intelligence and more people benefit
2:57from this content. But anyway, let's
2:59get started. What is an artificial
3:01intelligence agent? The first thing I
3:04want to tell you is that, unfortunately
3:06, even though this term is gaining more
3:08and more popularity, most of the people
3:10who work in this industry have an
3:11incentive—I think a perverse one—to
3:13make this much more complicated than it
3:15really is. So they tell you that an AI
3:20agent is written 100%in code, that it
3:21needs a brain to be able to function,
3:23and that they have to come in and
3:25configure it for you. And of course,
3:27there are different types of artificial
3:29intelligence agents. Without a doubt,
3:31there will be some that are much more
3:33sophisticated than others, but you can
3:35ask your ChatGPT or your Claude this,
3:36and you can ask them, "Hey, I heard the
3:38following in a YouTube video, tell me
3:40if it's true or not." And then you let
3:42me know what they tell you: an
3:43artificial intelligence agent is
3:45nothing more than a folder on a
3:46computer. That's all it is. I'm going
3:48to repeat it. An artificial
3:54intelligence agent is nothing more than
3:56a folder on a computer added to, of
3:58course, an AI model and something
4:00called a "harness," which is a computer
4:01program that allows us to combine the
4:03model with that folder so that the
4:05artificial intelligence doesn't just
4:07answer our questions, but—and this is
4:09key to understand—when we combine the
4:11folder I mentioned with instructions,
4:13connections, et cetera, the artificial
4:14intelligence model, and the harness
4:16that allows us to integrate it all. The
4:24artificial intelligence model no longer
4:26just answers our questions, but rather,
4:28we ask it a question or we ask it to do
4:30something, it goes and tries to do it,
4:32it looks at its own result, the result
4:34of its actions, and evaluates if that
4:36is enough or not to fulfill what we
4:38have entrusted it to do. And in case
4:44that’s not enough—and this is the
4:46key—it enters a sort of loop where it
4:48tries again to meet the objective,
4:50checks the result, and if it can't
4:51provide the deliverable we asked for
4:53with the current result, it keeps doing
4:55it, and doing it, and doing it until it
4:57finishes. And you might say, "That
5:01sounds very sophisticated, that sounds
5:03very complicated, but it isn't." Again,
5:05it’s just a folder, an artificial
5:07intelligence model, and a harness. And
5:10you’re probably wondering where you
5:11get that darn harness and how to
5:13integrate all of this. Well, don't
5:14worry, because we're going to see that
5:16right now. Well, fortunately, if you
5:19currently have a subscription to
5:21ChatGPT, Claude, Kimi, or Gemini, you
5:23already have access, just by having
5:25that subscription, to a harness that
5:27allows you to train and deploy
5:29artificial intelligence models. In this
5:35case, for this exercise, we are going
5:37to use Claude Code, which is a harness
5:39that can connect to an artificial
5:41intelligence model and that we can
5:42point to a folder to create an agent.
5:46But Claude Code is not the only harness
5:48we can use for this. We could use
5:51Claude Work, we could use ChatGPT Work,
5:53we could use Codex, we could use Kimi
5:56Code, Antigravity, and so on. All of
5:59these are harnesses that, when we
6:01connect them to an artificial
6:02intelligence model and point them to a
6:04folder containing instructions on what
6:06that agent should do, they turn into an
6:08artificial intelligence agent. And if
6:11you want to learn about Claude Work or
6:13ChatGPT, I’ll leave, I think in the
6:15comments or somewhere here on the
6:16screen, a couple of videos that
6:18Salomón and I made explaining how
6:20these tools work in just 20 minutes.
6:22But today, we are going to use Claude
6:24Code. And why Claude Code? Because
6:27unlike Claude Work or unlike ChatGPT
6:29Work, and similarly to how it works
6:31with Codex, Antigravity, or Kimi Code,
6:33these are harnesses that have fewer
6:35restrictions. They were initially built
6:41so that software developers could use
6:43agents to develop software, and
6:45consequently, they have fewer internet
6:47connection limitations, fewer
6:48limitations regarding the quantity and
6:50volume of files they can process, and
6:52so on. Claude Work and ChatGPT work
6:58within virtual environments, and that
7:00makes them difficult or sometimes a bit
7:02slower, making them less efficient for
7:04processing large amounts of files. In
7:08contrast, Codex, Claude Code, and
7:10Antigravity basically allow us to do
7:12whatever we want. The good thing is
7:15that Claude Code, Codex, and
7:17Antigravity, which used to be difficult
7:19tools to use because they lacked a
7:21user-friendly desktop application, are
7:23now very easy to access. We simply go
7:26to the desktop application for Claude,
7:28ChatGPT, or whichever one you prefer.
7:31I'm going to go to Claude's, select the
7:33tab up here that says Code, and that's
7:35it. Now we are inside Claude Code. We
7:40are inside our harness, which is what
7:41allows us to create that loop I
7:43mentioned—not just asking the model a
7:45question and getting an answer, but
7:46asking a question. It goes, tries many
7:50things, makes a number of connections,
7:52and then answers me. So, we already
7:54have the harness. Now we need the
7:56artificial intelligence model and the
7:57folder I mentioned. The artificial
7:59intelligence model in this harness,
8:01which is Claude Code, is basically
8:02already connected by default. You can
8:04see it down here on the right. Right
8:06now, I have a model selected called
8:08Claude 3.5 Sonnet. It is one of the
8:10most advanced models that exists today.
8:13It consumes tokens and resources like
8:15no one else in the world. And I have it
8:17set to max effort level. In other words
8:21, I am explicitly asking it to reason
8:22and perform that loop for as long as
8:24possible to give me the best answer.
8:29Normally, it is not necessary to run
8:31these models this way or with such deep
8:33reasoning, and usually, to build these
8:35types of agents, you don't need to use
8:37the most advanced model either. In fact
8:42, I suggest that for your tests, just
8:44to save tokens and also to measure
8:46whether this is enough or not and to
8:48gain speed, start with Sonnet 3.5. It
8:52is a mid-range model, let's say, but it
8:54is quite good. If you use ChatGPT, you
8:57can select—at the time of recording
9:00this video—one called GPT 4o, which
9:02is also quite good. And I normally use
9:06it in high or extra high for these
9:07types of workflows, but in this case,
9:09I'll leave it down here on high. Ready?
9:13So, the three legs of an artificial
9:15intelligence agent: a harness, a model,
9:17and we are missing the folder, right?
9:19Well, we already have the harness,
9:21which is Claude Code, we have the model
9:23configured, now we just need the folder
9:25. And in Claude Code, all we have to do
9:29is select the folder we want to work in
9:31using this icon you see here. I'm going
9:34to open a new folder here. I have a
9:39folder on my desktop called code, where
9:41I keep everything related to software
9:42development and coding. Uh, but you can
9:46do it wherever you like. and I create a
9:49new folder that I'm going to call
9:50prospecting agent. I'll explain why in
9:53a moment. You don't have to write it in
9:55lowercase, you don't have to put these
9:57uh dashes between the words, it's just
9:58a habit of mine. And we simply click
10:01open. Done. Perfect. We already have
10:04our harness. Let's say the software
10:07that wraps all of this and allows us,
10:09let's say, to use the artificial
10:10intelligence model iteratively. We have
10:13our artificial intelligence model
10:14configured and we have our folder. But
10:17as you just saw, that folder is new,
10:19it's empty. We have to fill it with
10:23something, and that something is a
10:25prompt that we have been perfecting
10:27over time, that we have been iterating
10:28on, that we took from different sources
10:30on the internet and adapted to our
10:32needs, and you can find it in the
10:33description of this video so you can
10:35download it, use it whenever you want,
10:37and it's the one you see here on the
10:38screen. It is a quite long prompt, you
10:45can read it, I'm going to explain it to
10:46you, but essentially this prompt is
10:48what explains to the artificial
10:50intelligence agent how it should behave
10:52. Without this prompt, we would give
10:57any instruction to the artificial
10:58intelligence agent and it would go and
11:00do its best to complete that task. But
11:06the next time we ask for that same task
11:07or a similar one, without this prompt
11:09we just saw, it will do the task again
11:11in a probably different way and the
11:13third time it does it, it will do it in
11:15a different way too. And this is
11:18problematic. Why? Because in the
11:22business context, we always need
11:23consistency, we need standardization.
11:28Imagine if we had a fast food business
11:30and we had a cook who always made a
11:31burger differently every time we asked
11:33for one. It's a good analogy for an
11:36artificial intelligence agent because
11:38you tell a cook, "Make a burger." And
11:41he goes, tries one thing, like checking
11:43the fridge to see if the meat is there;
11:45if it's not, he goes and checks
11:47somewhere else, he enters this loop I'm
11:49telling you about and, in the end, he
11:51does all these things; that is, he acts
11:53, reasons, acts again, reasons to
11:54fulfill the task we entrusted him with.
12:01But of course, ideally, that cook
12:03should be trained so that every time
12:05they make a burger, a hot dog, or fries
12:07, they follow the same recipe, which
12:09they can always improve, iterate on,
12:10and take notes on as they learn from
12:12practice, but stay consistent. And that
12:19is a bit of what we are teaching here.
12:21So, pay close attention. This prompt
12:28tells people that we want them to work
12:29in a structure we have called D.O.E.
12:31And again, we didn't invent this; we
12:33found it from various sources on the
12:35internet. Several people have talked
12:41about this, but D.O.E. basically stands
12:44for Directives, Orchestration, and
12:46Executables. And how does it work? What
12:52we are telling this person is, please,
12:54before starting within the folder we
12:55are working in, create two main folders
12:57. One folder called Directives and
13:03another called Executables. When I ask
13:07you for something the first time,
13:09please go and do it in the best way you
13:11see fit, according to the prompt and
13:12the instructions I gave you. But pay
13:17attention, document what you did, how
13:19you did it, and why you did it. In a
13:23plain text file. There is nothing
13:25special about this, it is just a plain
13:27text file. And put it in that folder we
13:30mentioned called Directives. And mind
13:39you, any buttons or tools you used or
13:41think would be useful to complete that
13:43task associated with that directive,
13:45please document them in scripts or
13:46computer code sequences so they always
13:48execute the same way. Again, this
13:57computer code is also literally just
13:59files with text. The thing is, when I
14:04put computer code into a standard
14:06computer program, it always produces
14:08the same result. And that is very
14:11important. And why is it very important
14:13? Because let's assume I tell this
14:20artificial intelligence agent that its
14:22task will be to take a mountain of
14:24documents, say 5,000 documents that I'm
14:26going to throw into a completely messy
14:28folder, and I want it to read all those
14:30documents, organize them according to
14:32—let's assume they are invoices—the
14:34vendor, the date, etc., and register
14:36all that information in a spreadsheet.
14:45And also to reorganize all the files,
14:47let's say in the folder in a certain
14:49way, for example, by receipt date, by
14:51vendor, whatever we want. If I have to
14:56repeat all those instructions every
14:58single time I use the agent, then the
15:00agent isn't very valuable; or at least,
15:02we run into the same problem many
15:04people have: sometimes, giving
15:05instructions to people takes more time
15:07than just doing the work yourself. But
15:12if we tell them, "Man, do it once, and
15:13for everything you do, write down the
15:15step-by-step of what you did and why
15:17you did it in these guidelines, and put
15:19it in this folder. And to create the
15:26Excel file, I mean, to fill out the
15:27Excel file or to move the files, rename
15:30them, and put them in folders, instead
15:32of doing it manually, run a script or a
15:34little computer program that you also
15:36write so it’s always done the same
15:38way; well, then I can trust that when I
15:40give a pile of 9,000 documents to the
15:42agent and hit play, it will come and
15:44read: 'Hey, how did Martín like me to
15:46do the facial review?'" folders, well,
15:51with these guidelines. Perfect. Very
15:53good. So, I'm going to do it like this
15:56again. Oh, and these guidelines tell me
15:58that when I have to rename the files, I
16:00should run this little computer program
16:02that is also here in this folder, so I
16:04go and run it. So, it will always
16:08follow the instructions and execute the
16:10actions in the same way. And watch out,
16:15it can have many guidelines and many
16:17executables, because in the same folder
16:19or for the same agent, we can delegate
16:21many tasks and responsibilities. So, to
16:26recap, D stands for directives, because
16:29as we ask the agent for things, we are
16:31asking it to document how it does them
16:33in a text file called directives, which
16:35it places in a directives folder. And
16:41every time it takes an action, we are
16:43asking it to take the action, but
16:44create a computer code so that the next
16:46time that action is taken, it is done
16:48exactly the same way. And you might be
16:51wondering, "And what does the O stand
16:53for?" Remember that O stands for
16:55orchestration. And orchestration
16:58literally means acting as an
16:59intermediary between the user's request
17:01, the directives, and the executables.
17:04And this role is fulfilled by the
17:06artificial intelligence agent. So we
17:10are saying, "Lord, when I ask you for
17:11something, act as an orchestrator to
17:13decide which of the directives we have
17:15been creating together apply, and given
17:17those directives, which executables or
17:19buttons should you press so that things
17:21always happen the same way?" So,
17:24directives, orchestrations, and
17:27execution. And something very important
17:31is that as he gives us results, as I
17:32ask him for something and he
17:34orchestrates or intermediates between
17:35the directives and the executables, he
17:37will present me with results. And when
17:41he presents those results to me, I can
17:43tell him as a user, "Hey, that's wrong.
17:45" I don't like that you use that font.
17:50I don't like that you put these colors
17:52on the cells that are negative. I don't
17:56like it when you use decimals in the
17:57cells, I have no idea what one might
17:59want. And part of what we are teaching
18:04him with this prompt is that when I
18:05give him that feedback, he comes and
18:07records a lesson here in the prompt we
18:09just gave you. He can keep editing it,
18:14but also modify his directives and
18:16modify his executables so that the next
18:18time I ask for exactly the same thing,
18:20he does it the way he was corrected to
18:22do it. I hope this is clear, you can
18:25watch it as many times as you want. It
18:27is fundamental to understand this, but
18:29well, now let's get to the practice. We
18:31are going to copy all this, this whole
18:33prompt, and we are literally going to
18:35go back to Cloud Code. Remember the
18:38harness, the folder where we are going
18:40to work, and the selected model. We
18:44paste it and basically we are telling
18:46him here, this is the way you are going
18:48to act and create the folder system
18:49inside the folder I just gave you. And
18:55something I like to do is to open the
18:57files here, which, as you can see, at
18:59this moment the folder we gave the
19:00agent is completely empty, but it is
19:02now going to be populated with some
19:04files and some folders because we are
19:06giving it the instruction in this
19:07prompt that, again, you can download
19:09below, to create that structure that
19:11will allow it to work consistently for
19:13us. And that's it. A few minutes later
19:19he does what he has to do and look that
19:21from one moment to the next the folder
19:23is full of files. In fact, look, I just
19:26opened the folder and here it is full
19:28of files. The prospecting agent folder
19:31we created a minute ago, which had
19:33absolutely nothing in it, now has some
19:36files, and be very careful because the
19:38files it has are very important, at
19:40least these ones you see here called
19:42agents.md, cloud.md, and gemini.md.
19:46Let's start with cloud.md. MD stands
19:50for markdown, which is simply plain
19:52text with some indicators, like these
19:55hashtags that tell computers what the
19:57hierarchy of the text is; for example,
19:59one hashtag means this is a title, and
20:01so on. But notice that this cloud.md
20:07file you see here, or this gemini.md
20:09file you see here, are exactly the same
20:11or have exactly the same text from the
20:13prompt we just gave it. I don't know if
20:18you noticed, but look, it starts with
20:20instructions for the agent, create a
20:22cloud.md file, and here it says
20:24instructions for the agent, create a
20:26cloud.md file, and everything else. Why
20:28do we have three identical files?
20:33Because it turns out that when you are
20:35working in Claude Code, or working in
20:37Claude Cowork, or working in ChatGPT's
20:39Codex, or in ChatGPT Work, or working
20:41in Google's Antigravity, which are all
20:43harnesses for artificial intelligence
20:45agents. When you start a new session,
20:53each of these harnesses is programmed,
20:55before answering you, to look if there
20:57is a file in the folder where you are
20:59working called cloud.md in the case of
21:01Claude, agents.md in the case of Codex
21:03or ChatGPT Work, or gemini.md in the
21:05case of Google Antigravity. And if they
21:12find a file literally named that way on
21:14each of these platforms, what they have
21:16to do, or what they do, excuse me, is
21:18go and read it first. So, the fact that
21:22these three files exist in this folder
21:24means that I, this artificial
21:26intelligence agent (which, again, the
21:28agent is composed of the harness, the
21:30model, and the folder), I will be able
21:33to replace the Claude Code harness, for
21:35example, with ChatGPT, I mean ChatGPT
21:37Work, or with Codex. And in that case,
21:41when I speak to the artificial
21:43intelligence agent again, this time
21:44with Codex, what it will open is
21:46agents.md. But if I remove Codex and
21:50put in Claude Code, it will open
21:52cloud.md. And if I remove Claude, the
21:54Claude Code harness, and put in Claude
21:56Cowork, it will read cloud.md again.
21:58And if I remove Cowork and put in
22:01Gemini, or Google Antigravity, excuse
22:03me, it will read Gemini.md. Regardless
22:07of which one it reads, they will all
22:08have the same thing, and the agent is
22:10trained to always keep them identical.
22:13This allows me to keep working with
22:15artificial intelligence agents, but
22:17change the harness and, of course, the
22:19model whenever I want. Now, look at
22:21what I explained to you. Remember,
22:24there is a folder called directives,
22:26which is completely empty, and a folder
22:28called execution, where these scripts
22:29or these buttons go so that things are
22:31always done exactly the same way. And
22:34there are also other files or other
22:36folders. This is a temporary folder
22:39that it creates to generate temporary
22:41files and delete them like a kind of
22:42notepad, and some files that contain,
22:44let's say, secret keys, things that
22:46cannot be shared. This is a security
22:50issue; the agent handles it on its own,
22:52but with this, we already have our
22:54agent configured because we have our
22:55harness, our artificial intelligence
22:57model, and our folder. That alone, even
23:01if the folder were empty, is enough to
23:03have an agent. But now we have that
23:08folder with a way of working, a work
23:10structure, a mental model that
23:12guarantees that the agent will keep
23:13learning, keep documenting, and keep
23:15getting better and better. And I swear,
23:20this is absolutely mind-blowing. Once
23:26you really start working within this
23:28same folder on the same problem or
23:29series of problems, it is incredible to
23:31see how the agent starts to respond to
23:33you in the same way and to respect
23:34those explicit or tacit agreements that
23:36you make with it. And with this done,
23:42the only thing we lack to be able to
23:44train an artificial intelligence agent
23:46—and I am sure you never imagined it
23:48was as simple as opening Cloud Code,
23:50choosing a model, creating a folder,
23:52and pasting a prompt—is that we are
23:54ready. The only thing we have to do now
23:57to train this artificial intelligence
23:58agent is to start asking it for things.
24:02I like to have a folder for each
24:03function of my company. So, I have a
24:08folder for digital marketing, I have a
24:10folder for legal matters, I have a
24:12folder for administrative affairs, I
24:14have a folder for content creation, and
24:16I work with each of those folders
24:18individually to create these directives
24:20and these executables, and I simply
24:21switch from one to the other. In fact,
24:27to make this a bit more visual, I even
24:29created this application called Pulpo,
24:31where I wanted to see it a bit more
24:33like a video game. And each of these
24:37little figures you see here is an
24:38artificial intelligence agent that
24:40helps me with different things, like an
24:42administrative agent, software
24:44development, content, a legal
24:45consultant, a marketing agent, and so
24:47on. And this looks very sophisticated,
24:51but what's really behind this is just
24:53what you already know: a harness, a
24:55folder with some directives and
24:56executables, and a lot of training work
24:59. If I, for example, go to my marketing
25:05agent, I can ask it for anything, like,
25:07for instance, "help me create a cold
25:09email campaign for Aztec." And if I
25:13were to send it, it already knows how
25:15to create cold email campaigns for
25:17Aztec. It goes and looks at its
25:18directives, looks at its executables,
25:20and creates it. I don't have to do much
25:22else. And if you want me to release
25:26Pulpo to the world so you can download
25:28it too and use it on your own computers
25:30, devices, or phones to train AI agents
25:32, leave a comment and like this video
25:33so I know you're really excited and
25:35interested in this topic. But you don't
25:40actually need Pulpo, you don't need
25:42anything that flashy or elegant. All
25:45you need is a harness, a model, and a
25:47folder with a way of working to be able
25:49to train your own artificial
25:51intelligence agents. And the next time
25:54you want to use this agent or any other
25:56agent, you just have to come to Cloud
25:58Code, click new, choose your agent's
25:59folder, and ask it for things. But
26:02anyway, here we are. I'm going to
26:06return here to the conversation I was
26:07having with my agent and now I'm going
26:09to start training it. And what am I
26:11going to do? I'm going to create an
26:13agent. I named it prospecting agent
26:21because what I want it to do is help me
26:23get leads or people who might be
26:25interested in my products or services
26:27at Aztec; help me find information
26:28about those leads, that is, help me
26:30find their LinkedIn profile, their
26:32email address, investigate their
26:34companies, etcetera; put all that
26:35information into an Excel file or a
26:37spreadsheet with certain columns; and
26:39after analyzing all the information for
26:41each of those leads, write a
26:43personalized email for that person
26:44offering my products or services. Oh,
26:54and to add a little more excitement,
26:56let's tell it to make a dashboard for
26:58us that shows the composition of all
26:59those leads it found. How many are CEOs
27:04, how many are marketing specialists?
27:07How many live in certain countries?
27:09Like a kind of dashboard that allows us
27:11to see that information. This is not
27:14necessary, but I want you to see how,
27:16as I train it, it always gives me the
27:17same result or always does the work the
27:19same way. What is it that we want to
27:22achieve? Before we start, there is
27:27something very important that we need,
27:29and that is giving our agent the
27:30ability to connect with potential tools
27:32on the internet that could be useful
27:34for completing its task. We could not
27:40do this and simply tell it, "Hey, go
27:41and figure it out." And it will go and
27:45figure it out, it will find leads
27:47somehow and it will see what it can do.
27:50Search for them on Google. I have no
27:54idea, but obviously to the extent that
27:56we have an idea of what it will need to
27:57do this automatically, then we must
27:59give it those tools so it can make use
28:01of them. And also, of course, we want
28:08to give it access to our accounts in
28:10those tools, in case the tools need, I
28:12don't know, paid accounts, need a
28:14subscription, or whatever. We at Astec
28:21maintain that the best way or the main
28:23way we like to do this is by using a
28:25tool called Composio. We have nothing
28:29to do with Composio; in fact, there are
28:31other similar tools. There is one
28:35called Make, but we love Composio
28:37because it allows our cloud code or our
28:39artificial intelligence agent, with a
28:41single connection, that of Composio, to
28:43connect to more than 1,000 tools that
28:45exist on the internet. All we have to
28:50do is go to composio.dev and this is
28:52completely free. Create an account. If
28:57for any reason it redirects you to this
28:59platform section, go to "for you" and
29:01come and connect our accounts just once
29:03. So, if we need, for example, for our
29:08agent to use our Gmail account or
29:10Google in general, there is a fabulous
29:12account called Google Suite and we can
29:14connect several of our accounts to it,
29:16and Google Suite gives it access to
29:18email, calendar, Drive, etc. Or there
29:21is a fabulous account called Apify,
29:23which is the one you see here, that
29:25allows our agent to extract information
29:27from thousands of websites, such as
29:29social networks—I'm talking about
29:31Instagram, LinkedIn, etc. It can
29:33extract information from Google Maps,
29:35etc. Or there is a very famous
29:37prospecting tool called Apollo. Or here
29:40we are talking about a prospecting
29:42agent. I have an Apollo account. These
29:44are relatively inexpensive accounts. In
29:48fact, there’s even a free plan, but
29:50you can connect Apollo so that the
29:52agent can go and search for prospects
29:53using Apollo. I’m only talking about
29:57prospecting again, but you will find
29:59all kinds of tools here. If you don't
30:02know which tools might be useful for
30:04you so your agent can complete the task
30:06you need it to. Just ask it, tell it, "
30:08Hey, I want you to do this." What are
30:11the tools you need? "And if you are in
30:13Composio, this is the best way to give
30:15it access. So, we have a full class on
30:18connectors where we teach you how to do
30:20this. We’ll leave it somewhere around
30:22here, or in the description, so you can
30:24go and watch it. But, simply put, to
30:27connect Composio with your agent, you
30:29go to the install section, choose Cloud
30:31Code, Codex, or ChatGPT, whatever it is
30:34, and it guides you on how to install
30:36your Composio account. I’ve already
30:39done it. So, that means my AI agent is
30:43now a harness with an AI model, with a
30:45folder that has a way of working, and
30:47now it has a way to connect with over
30:491,000 tools on the internet. See? This
30:54is crazy. Now, the only thing left is
30:58to train it so that, whenever I ask for
31:00something, it does things the same way.
31:03So, how do I do that? It’s the
31:05simplest thing in the world. I just
31:07turn on the microphone and use a tool
31:09called Astec Voice, which is ours,
31:11it’s 100%free, you can find it by
31:13going to the tools section of our page,
31:15and you can download and use it. And
31:18this allows me to dictate to the
31:19computer. This is one of the best
31:22things out there, and there are many
31:24tools like this on the market, most are
31:26paid, this one is completely free, and
31:27we give it to you just for being part
31:29of our community. And what we are going
31:33to do is tell it in normal Spanish,
31:35like any friend, what we want. I want
31:38you to please help me create a workflow
31:39where, when I tell you I want to sell a
31:41product for my company, you help me go
31:43through it, ask me questions, and
31:44interview me about things I’m
31:45probably not seeing. Based on the
31:47information I give you, use Apollo to
31:48extract a list of 100 potential leads,
31:50and then I want you to look for the
31:52information, do an internet search for
31:53each of the companies where those leads
31:55currently work, and also extract all
31:57the information from their LinkedIn
31:58profiles, including, well, their
32:00profile information and any possible
32:01posts they may have made. And I want
32:03you to take all that information and
32:05put it into an Excel file, or Google
32:06Sheets, better, because I use Google
32:08Sheets. And that Google Sheets file
32:10must have columns for: first name, last
32:12name, the company they work for, their
32:14job title; it must have their LinkedIn
32:16profile URL, a summary of the company,
32:18and the raw information from their
32:20entire LinkedIn profile and their posts
32:22, let's say, their last five posts. And
32:24in the last column, I want you to
32:26create a hyper-personalized email for
32:27each of those leads offering them the
32:29product we discussed, but I want you to
32:31write that hyper-personalized email in
32:33a casual way, I want it to seem like—
32:34and you use the information we just
32:36extracted. So, I want it to look like
32:37an email I’m receiving from someone I
32:39might have met at a conference or a
32:40coffee shop, to generate that impact,
32:41that impression, and say something
32:42related to the information we have
32:44about him; something like:" Hey Gabriel
32:45, I was looking at your LinkedIn
32:46profile and saw your post about the
32:47conference you gave in Guadalajara. "
32:49It's incredible how easily you explain
32:50the intersection between artificial
32:51intelligence and marketing. I'm writing
32:53to you because I kept looking at your
32:53profile and it occurred to me that this
32:54might interest you, something very
32:55casual like that, and then you hit them
32:56with the pitch. And I want you to do it
32:58this way, don't use, like, excessive
32:59punctuation marks. It’s like a very
33:01casual email, 100%text that we are
33:02writing to the person. And then you
33:04give them the pitch and make them an
33:05offer. And on top of that, when you
33:07finish 100%of that work, I want you to
33:09create a dashboard where you can show
33:10me, or I can interactively see, let's
33:12say, where the leads are located, their
33:14emails, where they work, etc., and I
33:16can choose lead by lead each one of
33:18those or see all the information lead
33:20by lead. Uh, please, ask me every
33:22question you think is uh necessary to
33:23fully understand my situation and to
33:25give me the best possible answer. And
33:27folks, that is the magic of Astec Voice
33:29, right? These transcription tools. I
33:32mean, how long would it have taken me
33:34to write all of this? Really, having to
33:37type really limits one's ability to
33:39communicate with computers, especially
33:40with artificial intelligence, but these
33:42tools allow us to give it all the
33:44context possible. And pay attention,
33:47that question I ask at the end, this,"
33:49please, ask me every question you think
33:51is necessary to fully understand my
33:53situation and give me the best possible
33:55answer, "is one of the most important
33:57things that exists in artificial
33:59intelligence. Get used to doing this.
34:03Don't do prompt engineering, don't get
34:05into prompting courses, that's useless.
34:10Turn on the microphone, give it all the
34:12context possible and tell it to ask you
34:14questions, and you'll see that it will
34:15ask us very interesting questions about
34:17what it needs for what we are looking
34:19for and what it needs to help us better
34:20. Look, the first thing it does after
34:25thinking for a while is that it tells
34:27me," Hey, before starting, one
34:29observation, these are actually two
34:30linked projects, the pipeline with the
34:32trigger phrase and the dashboard. "" I
34:35prefer to design and build number one
34:36first with your own spec and design the
34:38dashboard later. "Let's tell it perfect
34:40, sounds good to me, do it that way, I
34:42have no problem. And again, look how
34:45interesting it is when you just start
34:46talking to the computer. It is, it's
34:49amazing how one can work nowadays with
34:52these. And I want to come back down to
34:54what we are doing. Remember, what we
34:58are doing is we took a harness, we took
35:00an artificial intelligence model and a
35:02folder. That, together, makes an
35:09artificial intelligence agent, and we
35:10gave the folder a way of working that
35:12you will be able to download here, uh,
35:14that will allow us to make sure its
35:16results or the product it generates is
35:17always consistent, because it will be
35:19noting and learning as we give it
35:21feedback. We also connected the harness
35:27to Composio and again, if you don't
35:29know about connectors or want to know
35:31more about connectors, we have a
35:32completely free class. In fact, that
35:34class is an exclusive class from our
35:36artificial intelligence course that we
35:38put here on YouTube. So, go and watch
35:40it. And it's already asking me some
35:41additional questions, which is how do
35:43you prefer to obtain the information
35:44from LinkedIn. So it tells me Apollo
35:46plus public web API for LinkedIn
35:48enrichment. Remember that we already
35:53gave it Composio, so we'll just tell it
35:55, you have access to Composio, please
35:57use it, and there you will find both
35:59Apollo and Apify to extract the
36:01information from LinkedIn and the
36:03internet search; go solve it. Notice
36:10how it keeps asking me questions, as I
36:12asked it to, to try and refine my idea
36:14or what I want. And often, by answering
36:19these questions, one realizes things
36:21they hadn't seen before, I mean, things
36:23they hadn't noticed. And now we just
36:27have to wait for it to either ask us
36:28new questions or start building its
36:30ability to create these cold emails.
36:40One thing: in this video, we will only
36:41go as far as writing the emails as a
36:43way to demonstrate the consistency we
36:45can achieve by training these agents,
36:47but we won't learn how to create the
36:49cold email campaign and send these
36:51emails to get leads and answer them,
36:53etc. If you want a video on that and on
36:55how to create this whole framework
36:56using AI agents, but to be able to
36:58effectively send thousands of emails a
37:00month and get responses and generate
37:02leads, leave us a comment, because if
37:04there are enough comments saying," We
37:06want the Cold Gmail or cold email video
37:08, "we'll do it. It is absolutely
37:16incredible the number of leads one can
37:18generate with these well-configured
37:20campaigns. Notice that just by me
37:22mentioning Composio, since it already
37:24had it installed, it says perfect, I
37:26know how to use it, and it goes and
37:28uses it. And something interesting is
37:30that you can see it literally goes and
37:32solves it. Remember how I told you the
37:36harness creates this loop that allows
37:38the AI model to think, use tools, look
37:40at the response, think again, and use
37:42more tools? See how it thinks," Hey,
37:48the official LinkedIn connector doesn't
37:50work for this, but through this tool
37:52called Apify I can use this API that
37:54lets me bring the information back and
37:56I can also use this other one for posts
37:58, etc. "And if you don't have an Apify
38:00account, you can create one here
38:02completely for free. They give you, I
38:08think, $ 5 to use for data extraction,
38:10which is plenty to at least get started
38:12, and eventually, when you spend those
38:14$ 5, you'll have to pay, but by then it
38:16will probably be worth it. Although if
38:20you want cheaper options to extract
38:22information from social networks, uh,
38:24from websites, etc., you know: join our
38:26community, where we share this
38:27information every day, and we'll leave
38:29the link in the comments; or leave us a
38:31comment, uh, asking for this
38:33information and, I don't know, maybe
38:35we'll make a video about it. Something
38:38very interesting is that it just
38:39stopped here to ask me a question and
38:41says," Hey, before we continue, I
38:43noticed something: you gave me that
38:44list of columns you want in your Google
38:46Sheet, but you forgot to tell me the
38:48email. "And the email is fundamental
38:50for being able to send emails. So it
38:52tells me," Hey, should I include the
38:54email? "And I'm like," Sure, add it. "
38:56Perfect. And it's telling me how it
38:59plans to structure the directives and
39:00execution. I'm going to tell it," You
39:02make the decisions. That sounds good to
39:04me. Uh, thanks for noticing the column.
39:07Add the email one and now let's wait
39:09for it to create the directives and
39:11executables, configure itself, and
39:13let's see the first result. Well, this
39:15took a while. Uh, in fact, I even went,
39:18ate, and came back. The guy went,
39:21thought, did everything, and reached
39:22this point where it asked me a question
39:24that I love that it asked, because I
39:26was going to talk to you about this
39:28later, but it beat me to it and says: "
39:29Hey, how do you want me to execute the
39:31eight tasks in the plan?". Do you want
39:36me to deploy sub-agents to do the work
39:38in parallel or do you want me to do it
39:39all one after another. And obviously,
39:44this is one of the features that agents
39:46or artificial intelligence harnesses
39:48have, of artificial intelligence agents
39:50, and that is that they can deploy
39:52sub-agents. What does that mean? That
39:55they can create mini-instances of
39:57artificial intelligence agents to do,
40:00for example, if there are 100 emails we
40:02are going to process, to do 25, 25, all
40:04at the same time or, for example, to
40:07fulfill different roles. If you want to
40:10know a little bit more about sub-agents
40:13, watch our video on Chat GPT Work,
40:14where Salomón explains this very well,
40:17how it works. But of course, almost
40:19always I'm going to want to do it with
40:21sub-agents because then the work is
40:22done in parallel. Obviously, there will
40:25be situations where I want one thing
40:27done first, then another, and then
40:29another, but when I have a lot of
40:30similar work to do, it's good to
40:32separate it into sub-agents. So, I'm
40:34going to tell it to effectively use
40:36sub-agents. And look, it’s been
40:38thinking for 44 minutes. And you might
40:40think, "Well, if it's going to take 44
40:42minutes every time I ask it to do this,
40:44then it’s not worth it, it doesn’t
40:46make sense." But this is only the first
40:48time, this is only while it configures
40:50itself, creates the directives, the
40:52executables, and so on. After this,
40:56I’ll click once and, in minutes, it
40:58will be able to do the work
41:00consistently; because it will have
41:01already done all the work of seeing
41:03where to get the information, how to
41:05process it, and which buttons or
41:06settings to always run so the work is
41:08done the same way. And that's it. After
41:13a while, it finished and tells me it's
41:15ready. Notice, something I want to
41:18highlight is that here in the folder,
41:20there’s a directive that, if I open
41:22it, is literally what I told you,
41:24it’s just plain text, and some
41:25executables remained. So whenever I
41:30open this folder with a harness and an
41:33AI model, because of how this is
41:35structured, it will come, it will read
41:37agents.md or cloud.md or geminite.md.
41:41They are the same. That will tell it: "
41:45You are the orchestrator between the
41:47request the user makes, which is 'help
41:49me create emails for 100 people in my
41:51company to sell a product'," and it
41:53will come, read this directive, the
41:54only one there is for now, and this
41:56directive will guide it to a number of
41:58executables so that it performs the
42:00extraction or each of those actions,
42:02always in the same way. Let's do an
42:06exercise. So, let's tell this guy the
42:09following. Alright, I want you to first
42:11go to my website, which is ascllab.co,
42:12so you can see what we do. And I want
42:14you to help me create a campaign where
42:16we are going to sell customer service
42:17chatbots. Uh, chatbots, custom-made
42:18customer service artificial
42:20intelligences, uh, with, uh, knowledge
42:21bases with the ability to query
42:22knowledge bases, uh, so they don't
42:24hallucinate, uh, that also connect to
42:25the clients 'CRM and have an interface
42:27that allows their agents, or their
42:28human agents, uh, to take over the
42:30conversation or the AI agent to assign
42:31the conversation. In, in general terms,
42:33in general terms, I think the sales
42:35angle...These are all like, the, the
42:37features, but the sales angle is, uh,
42:39is, uh: sell more, retain more. To your
42:41clients, uh, with an AI customer
42:42service solution that works 24 hours a
42:44day, uh, doesn't complain, and works,
42:46uh, perfectly and is perfectly
42:47infinitely scalable. Something like
42:49that. I think you could sell it to any
42:51industry that, you know, has a high
42:52volume of customer service requests. I
42:54would focus for now on mid-sized
42:55companies. I think it's easier in those
42:57companies to reach the decision-maker.
42:58Uh, and well, I think that, uh, will
43:00help me find these leads. If you want,
43:03give me a test first of three leads
43:04that quickly meet these characteristics
43:06. Uh, and in fact, I think that should
43:08be the way we work from now on, and
43:10after that, uh, if I agree with what
43:11you're showing me, then on to the 100s.
43:13And keep in mind, what do I want to
43:15highlight? We took all this time, all
43:18this effort training the agent to
43:20create this directive and these
43:22executables. It means that it is
43:25already going to do its job this way.
43:27Whether we like that way or not, we
43:29don't know yet. It's the first time we
43:32are testing the agent with this prompt
43:33we just gave it, but when it shows us
43:35the result, we are going to tell it, "
43:37Hey, I don't like those leads. I don't
43:39think you're understanding the task
43:41well. This is what I want." And it will
43:43modify those directives and executables
43:45so that the next time it shows us
43:47something, if we tell it we like it
43:48that way, it will keep them like that.
43:51That's the first thing. And second,
43:53this whole process we just did, we only
43:55have to do it once. From now on, all I
44:00have to do is ask the agent for leads,
44:01see if it does a good job or not, and
44:03give it feedback, and the agent will
44:05keep giving me consistent results all
44:07the time. But also, and finally,
44:14another advantage of this is that,
44:16again, since an AI agent is a harness
44:18plus an AI model plus a folder, if I
44:20take this folder and go to another
44:22harness, like for example, Codex or
44:24ChatGPT Work, I can open that folder in
44:26that harness and again, without doing
44:28anything else. All I have to do is tell
44:33it, "Help me find 100 leads." And that
44:36harness will do it exactly the same way
44:37it did right now with Claude. Maybe a
44:40little bit better because the model is
44:42a little bit better. Maybe a little bit
44:43more affordable because the model is a
44:45bit more affordable. But look at how
44:47very portable this is. And when I say
44:48carry the folder, I’ll show you an
44:50example of how I’d do it with Codex.
44:52Here I am using the desktop application
44:54for Codex or Chat GPTIN. And look,
44:57I’m in the Codex section. Up here I
45:00already have the harness, I already
45:02have the AI model, which in this case
45:04is 5.6 sol light. I can also configure
45:07the effort here, which in this case is
45:09light. And the only thing I’m missing
45:13to have my lead and email generation
45:15agent is to bring the folder. So, I’m
45:19going to tell it I’m going to bring
45:20an existing folder and you already know
45:22, we’re going to go to documents,
45:24code, prospecting agent, which is where
45:26I have my folder with my directives and
45:28everything, and I’m going to open it.
45:30And look, I’m going to ask it, what
45:32can you do? And notice, this is while
45:35the other agent is running. It
45:37doesn’t matter. It’s a harness, an
45:41AI model, and a folder that has some
45:43directives and some executables. And
45:46look, magically in another application,
45:49the agent we built appears. I can help
45:52you build and operate this end-to-end
45:53prospecting system. In this workspace
45:57specifically, I can do the following:
45:59the directives, the execution, creating
46:01the deliverable, etc. See how the agent
46:03, by having built this folder, becomes
46:06portable and we can execute it here, we
46:08can execute it in a solution like, for
46:10example, Pulpo, we can execute it on a
46:12computer, in the cloud, on our friends'
46:15computers, wherever we want. This is
46:18the magic of what we are seeing here.
46:23It’s precisely this, how the
46:24instructions and those agent
46:26capabilities live in that folder, and
46:27since it’s nothing more than a folder
46:29with some plain text files, we can take
46:31it wherever we want and essentially
46:33change providers whenever it suits us
46:34or change environments whenever it
46:36suits us. We are not tied to a single
46:40provider. And well, finally after a
46:42good while it comes back and tells me
46:44that the leads it proposes are these.
46:46Silvia Galloso Velázquez from
46:48Profuturo AFP. Nicolás Ceballos and
46:51Miguel Ramos. There is something that I
46:54don’t like, for example, and I love
46:55that these things happen because that
46:57is how one trains the agent. So, I tell
46:59it something like, no, I think what you
47:01need to do is based on what I asked you
47:03for. Help me first by proposing some,
47:05uh, new company characteristics. How
47:07are you going to, what filters are you
47:08going to use? For example, I don't know
47:10, where are you going to look for
47:11clients, in which industries, what type
47:12of people are you looking for, I mean,
47:13CEOs or customer service heads, let's
47:14agree on that. Then, uh, quickly run a
47:16search for those people on Apollo, uh,
47:18and show them to me. And if we see that
47:19it works, then you do the whole
47:20exercise. Uh, afterwards, but first
47:22let's try to agree on the filters; uh,
47:23then, let's see that those filters
47:24actually work and, when I say: "Hey,
47:26the people you're showing me there
47:27really do seem like the type of person
47:28I want to look for," then you go and
47:30create the, the, uh, emails. And here
47:31we go again. So, now he says, "Hey, who
47:33are we going to target?" Uh, I would
47:37say the CEO, the founder, and the
47:39operations head. Well, I also like the
47:43customer service one, I like them all.
47:45And I had told him medium-sized
47:47companies, so I'll say 11 to 200
47:50employees, he tells me Colombia, Mexico
47:52, Argentina, Chile, and Peru, I like
47:55those. Yes, those five countries. Uh,
47:57filtered by specific industries or do
47:59we leave the search open? I would say a
48:01specific list. So he says retail or
48:03e-commerce, telecom, fintech banking,
48:06insurance, healthcare, SaaS, tourism,
48:08delivery, logistics with a high typical
48:11volume of customer service requests.
48:13That sounds good to me. And again, I
48:15want to be clear about what we are
48:16doing here. I am simply trying to work
48:21with him, building a way of working
48:22between him and me. So, he told me, "
48:26Hey, I have those three for you. That's
48:28not really what I'm looking for. What I
48:31want is for you, I tell you, I want to
48:33sell this more or less to these people,
48:35and for you to answer me." Well, then I
48:38suggest we use these filters to find
48:40those people. I'm saying this, let's
48:43say, after he gave me a first result
48:44that I didn't like, and it occurred to
48:46me that this could be a good way to
48:48work. If I have no idea what could be a
48:50good way to work, I'm going to tell him
48:52, "Give me ideas on how we can work.""
48:56And we will permanently be, let's say,
48:58or for quite a while we will be going
49:00back and forth, back and forth until we
49:01find a good way of doing things that he
49:03will go and document in the guidelines.
49:08Now, I really like, when I find
49:10something I like, to tell him:" Hey,
49:12document this, so that the next time we
49:14work, you do it this way. "And so he
49:18will explicitly receive an instruction
49:20from me to go and modify the guidelines
49:22and the executables so that this
49:24process is always done the same way.
49:26Now he gives me a new table, he tells
49:28me," Hey, I'm going to give you 15
49:30leads so that you can tell me if you
49:32like these. "So he tells me Manuel,
49:35Customer Experience Manager at XKNET
49:37Cancun, uh Gabriel also Customer
49:39Experience Manager at X28 Alarms and so
49:42on, but this doesn't give me much
49:44information. So, again, one doesn't
49:47simply settle for that, one tells him,"
49:50No, one second. In order for me to know
49:53if these are good leads, I need you to
49:55give me the full name, the position,
49:57the company; but also give me, like a,
49:59let me click on the company name to go
50:00to that company's page and see and see
50:02if it's the type of company I want to
50:04sell to. So, let's tell him exactly
50:07that. I want the workflow to always be
50:09the same. I tell you, I want to sell
50:10something. You ask me some questions,
50:12you make a proposal of what will be,
50:13like, the filters we are going to use
50:14to find those people. And mind you, I
50:16want you to show me a table like the
50:17one you are showing me. I like it, but
50:18show me the first name, the last name,
50:19show me uh the uh the position and show
50:21me their LinkedIn profile, with with a
50:22link for me to go to the LinkedIn
50:23profile and look at it. And show me the
50:25company, but for the company name also
50:26create a link so I can open the company
50:27page and know if it is a company I want
50:29to sell to. And well, a few minutes
50:31later he gives me exactly what I wanted
50:33and I just verified that he modified
50:34the guidelines. So now he tells me, "
50:37Well, Manuel, uh, last name Tun
50:39Contreras, Customer Experience Manager
50:41and here I can, well, this one isn't
50:43right because he didn't send me the
50:44company link. Let's look at Gabriel
50:47Fernández, let's look at his LinkedIn
50:49profile, he's from Argentina, uh, and
50:51works at X28 Alarms. Uh, let's tell him
50:54we want to open it in the browser. A
50:57good company, well, I do believe they
50:59might be interested in a chatbot.
51:02Actively. Could be. Is that okay? Uh,
51:05let's think about, uh, Andrea López
51:08Marroquín, customer experience manager
51:10. Let's look at the profile and the
51:13page is Aprendamos, okay? From Bogotá,
51:16or well, she lives in Bogotá. Human to
51:19human hop, the company is Aprendamos
51:21Academia, an education academy. My
51:25internet seems a bit slow right now,
51:27but yes, well, judging by those three I
51:29see, they work. So, I think we already
51:34have the leads, or rather, more than
51:35just having the leads, we are in
51:37agreement with our agent that we just
51:39built that this is the workflow we want
51:40to follow with it. Ready? So, I'm going
51:44to tell it that's very good. Uh, they
51:45seem to be fine. Uh, I think it
51:47documents that this is indeed the way
51:48we want to work. And now the only thing
51:50left is that, uh, next time I ask you
51:51for a list of leads, uh, or, or that I
51:53want to sell a product, that you always
51:55follow this process. Uh, go ahead and
51:56let's not do the 100 this time. Uh,
51:58simply, uh, extract all the information
51:59for these 15 and generate the emails
52:01for these 15 and leave me the Google
52:02Sheet just as an experiment. Do not
52:03change the directives for this. I want
52:05the directives to stay at 100 at a time
52:06, uh, but, uh, or well, we can
52:07eventually change them, but for now
52:09it's 100, but, uh, let's process these
52:1015 and show me the table. Oh, and
52:12another thing I always try to tell them
52:13is to also review the directives and
52:19executables so that most of the process
52:21, since we agree at least up to this
52:23point, is done programmatically and
52:25quickly. without much intervention on
52:31your part, uh, and if you need to
52:32parallelize the work, whether with
52:34sub-agents or, uh, in some other way,
52:36do it so the experience is faster. So,
52:42I really try, whenever I'm building
52:44these agents, to tell them:" Hey,
52:46everything you can move from you
52:48reasoning and thinking to executable
52:50files that run with a single click, do
52:52it that way, because, first, the
52:54results will be much more consistent,
52:56but also, the experience will be much
52:58faster. "If it has to take each one of
53:03the leads and check if it works for it
53:05or not, etc., it's going to take a long
53:06time doing it. One alternative, as I
53:09just mentioned, is to deploy sub-agents
53:11and parallelize the work. But another
53:14is to put everything you can into
53:16executables, so you don't have to think
53:18every time you need to do it again;
53:20just do it, because executables run
53:21immediately, instantly. And that, of
53:25course, is something that benefits us
53:26to the extent that we get answers much
53:28faster. And look, this is very
53:30interesting about the sub-agents. See
53:35how here, for example, a sub-agent
53:37deployed by our main agent is sending a
53:39message to the main agent, notifying it
53:41of a pattern that is worth the main
53:43agent knowing about in case it appears
53:44in other batches. And it's that the
53:50three leads had a job title or company
53:51different from the company or job title
53:53fields. Interesting. So, see how they
53:57collaborate with each other to solve
53:59and resolve, let's say, these possible
54:01issues they might encounter along the
54:03way. And that's it, after running
54:06through the leads, the 15 leads we
54:08mentioned, it leaves me a document here
54:10in Google Docs, just as I requested.
54:12Why? Because remember: through Composio
54:16, it has access to my Google Workspace
54:19accounts, and here is exactly what I
54:21asked for: the first name, last name,
54:24company, job title, email, LinkedIn URL
54:26, company summary, LinkedIn profile,
54:28recent posts, and the personalized
54:31email. In this case, the recent posts
54:37are taking up a huge amount of the
54:38spreadsheet's height, but what we can
54:40do is just ask it to align everything
54:42to the top so we can see it. So, Mario
54:47Hernández, who works at Toyota CoAPA,
54:50is the customer service manager. This
54:54is his email, this is his LinkedIn
54:56profile. Let's see if it's correct. It
54:58is. And it writes to him in the
55:00following way. I am fascinated by how
55:02it writes to Mario. Great post you
55:04shared on how in after-sales service,
55:06trust drops more due to lack of
55:07communication than wait times. With
55:09your nearly 10 years managing customer
55:11service at Toyota CoAPA, you surely see
55:13that every day. I'm writing to you
55:15because we made something just for that
55:17, an AI customer service chatbot that
55:19only responds based on actual agency
55:20knowledge, blah, blah, blah, blah, blah
55:22, blah. And it does this for all the
55:24leads. This is truly incredible. And
55:27it's all set. The agent is now capable
55:31of performing this same process in this
55:33same way with any harness and any
55:35artificial intelligence model because
55:36the directives and instructions on how
55:38to work are already in the folder. And
55:42for that, well, let's see it. Let's try
55:44to open that same folder. Now in Codex,
55:47which is another harness, the one we
55:48had already seen here. Let's just start
55:51a new conversation. We are in the
55:56prospecting agent and let's tell it,
55:58and since in this case the leads we
56:00just processed are in the temporary
56:01folder that it created here, you can
56:03see it here in the temporary folder.
56:08I'm just going to tell it to continue
56:10the conversation in Codex. So I am
56:15going to tell it, we just processed
56:17some leads with another harness and you
56:19can find that information in the
56:20temporary folder. Please help me review
56:25that information and let me know when
56:26you are ready. Perfect. I want you to
56:32tell me if you see the URL of the
56:34Google Sheets where that information
56:36was left, and you won't always find,
56:37let's say, the files in the temporary
56:39folder because the agent is instructed
56:41to delete them once we finish. Uh, but
56:46let's say that eventually you could ask
56:47it to save this information. It tells
56:50me there that it doesn't find it. So,
56:53uh, I'm just going to tell it that this
56:55is the link. This is the link. Please
56:58use Composio to access it. I want you
57:00to remove the em dashes, or those long
57:01dashes that give away that they are
57:02emails written by artificial
57:03intelligence, from all the emails and
57:04replace them with something, let's say,
57:06appropriate. Also, modify the
57:07directives so that you no longer use
57:08these long dashes or em dashes and from
57:09now on use language that doesn't look
57:11like it was written by artificial
57:12intelligence. The tone of the emails
57:13turned out quite well. What didn't turn
57:15out well was that punctuation mark that
57:16I don't like. And, you see, once you
57:20have this folder, it's like training;
57:22it's like a memory that you insert into
57:24the harness, into the artificial
57:26intelligence model, so that it can, in
57:28a second, see how it already knows what
57:30it has to do, how it has to do it, how
57:32it has to behave, etc. And we are going
57:34to let it do this in a second, but
57:36right after we are going to ask it to
57:37finish with this new Codex harness by
57:39creating the dashboard we said we
57:41wanted to create so that a dashboard is
57:43always created when we extract leads.
57:49Since our agent is already built in 2
57:51minutes, it went, found everything,
57:53caught up, and updated the Google Sheet
57:55. Now the email for Mario doesn't have
57:59those lines that bother me, and I
58:01assume the others are uh, the same
58:03without those lines. Perfect. But now
58:09we just have to show you how, in this
58:11new harness—let's say in this new
58:13environment where our same agent is
58:15living, because it's the folder, it's
58:17like its memory, its instructions, the
58:19way it works and communicates with the
58:21world—uh, but now using another AI
58:23model and another harness, we can keep
58:25working with the agent workflow. And
58:29for that, I'm going to tell it:" Now I
58:31want you to create a new directive or
58:33modify the directives and executables,
58:35so that whenever you give me the result
58:36of a lead generation, their information
58:38, and email drafts, you present me with
58:40both the Google Sheet and a board or
58:42dashboard where I can explore, let's
58:44say, all those people, the companies
58:45they work for, locations, job titles,
58:47and so on. "Make it uh very well
58:49designed, I want it to look uh really,
58:50really good, and yes, I want you to
58:52give me one of these dashboards with
58:54the information from that specific run
58:56every single time. Don't publish it
58:57anywhere, just give it to me here on
58:59the computer so I can see it, and
59:00obviously, make sure to modify the
59:01directives and executables so it always
59:03happens the same way. We hit enter and
59:04wait a minute, and that's it; it's
59:06already processed and gives me a link
59:08to open the dashboard. And look,
59:12obviously I would make some small
59:13changes, etc., but here it already has
59:15all the people it was able to extract.
59:21Let's do an exercise just keeping in
59:23mind that this already exists and is
59:25configured, and we have our agent
59:26completely ready to always do the task
59:28the same way. What we're going to do is
59:32create a completely fresh conversation.
59:36However, again, we have the harness, we
59:39have the AI model, and the folder with
59:41our directives and executables. And
59:45we'll simply tell it: please extract
59:47another 15 leads with the same
59:48characteristics as the last run, and do
59:50everything you know how to do. And
59:53that's it. A few minutes later, 5
59:56minutes to be precise, this guy has
59:58already done all the work. We have a
1:00:03new Google Sheet with 15 leads duly
1:00:05identified and with the emails written.
1:00:11Here, let's say there is something I
1:00:12would have to teach it to stop
1:00:14including and put in a directive, which
1:00:15is the subject line. But it also gives
1:00:23us a dashboard now that automatically,
1:00:25using another harness, the same AI
1:00:27agent, but using another harness and
1:00:28another AI model, it created a similar
1:00:30dashboard because, again, it's done
1:00:32with an executable that now has the new
1:00:34leads. Look, Carla Valle Salgado. It's
1:00:43Carla Valle Salgado, the one from this
1:00:45new batch or elite group. This is
1:00:48incredible. It's a folder we built
1:00:52slowly, simply by giving it
1:00:53instructions and providing feedback
1:00:55regarding what we liked and what we
1:00:57didn't like. And now we have an AI
1:01:01agent. Every time we need to do this
1:01:05work or any other work we train it for,
1:01:07we open a harness, choose an AI model,
1:01:09point it to that folder, and ask for
1:01:11the work to be done, and it magically
1:01:13gets done. So, what is life like with
1:01:17AI agents? You basically have a folder
1:01:21on your computer or in the cloud with
1:01:23subfolders. Each of those subfolders is
1:01:29like a card you can insert into your
1:01:31harness and your AI model so it adopts
1:01:33a certain personality or learns certain
1:01:35skills. So, you can have a folder that
1:01:41helps you clear your email every
1:01:43morning, and you train it to read all
1:01:45your emails; and you receive thousands
1:01:47of emails, and it separates them and
1:01:49answers the ones it should answer,
1:01:50notifies you of the ones it needs to,
1:01:52and so on, and you simply open Codex,
1:01:54Cloud Code, Cowork, or ChatGPT Work in
1:01:56the morning, point it to the folder,
1:01:58and tell it:" Help me clear my email. "
1:02:02" And it goes and clears it. "Or if you
1:02:04need one to send messages to your
1:02:06prospects to update the CRM, you train
1:02:08it in another folder and tell it to
1:02:10update the CRM. You can have a single
1:02:14folder that has multiple skills. I like
1:02:17to have, as I showed you just now,
1:02:19let's say folders that encompass entire
1:02:21functions. I don't have one agent that
1:02:23writes contracts and another that
1:02:25reviews contracts. I have a legal agent
1:02:28that I've taught to do all these things
1:02:30, and it's an exercise of iterating,
1:02:32improving, asking the agent for things,
1:02:34seeing the result, and asking it to
1:02:36correct itself. Fortunately, very
1:02:42quickly you end up with an agent that
1:02:44does things almost perfectly, and
1:02:46eventually, you end up with an army of
1:02:48virtual employees that help us be much
1:02:49more efficient because they work for us
1:02:51and learn without complaining. If you
1:02:56liked this video, put it into practice.
1:02:58Don't just stick with the theory, put
1:03:00it into practice because this
1:03:02transforms, it transforms lives. And if
1:03:06you want to know more about how to
1:03:08deploy or maintain these agents on your
1:03:10computers, on cloud computers, if you
1:03:12eventually want access to Pulpo or want
1:03:14me to make a video on how Pulpo works
1:03:16and how I manage AI agents in my
1:03:18day-to-day life, leave us a comment.
1:03:24And don't forget to subscribe to our
1:03:26channel if you want to learn more about
1:03:28automation, artificial intelligence,
1:03:30vibe coding, how to build technology
1:03:32products, and, in general, how to be
1:03:34more efficient and ride this wave that
1:03:36is changing the world and will
1:03:37definitely transform the economy and
1:03:39how we relate to it. I thank you very
1:03:43much, I hope you learned something,
1:03:44leave your questions in the comments.
1:03:46See you soon.

About this transcript

This page contains the full transcript of Crea agentes de IA en Claude que trabajan por ti en 60 minutos. by Aztec Lab, generated from the public captions YouTube serves with the video. The transcript has 11,448 words across 1,646 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.