YT YouTube2Text

YouTube2Text → Transcripts → You’re Not Behind (Yet): How to Build Your First AI Agent (Full Guide)

You’re Not Behind (Yet): How to Build Your First AI Agent (Full Guide) — Transcript

by Dan Martell · 5,215 words · 696 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00I just read a study that by 2030, AI is
0:02going to create 170 million new jobs,
0:05but they won't be jobs where you just
0:07sit there and chat with AI. They'll be
0:08jobs where you build AI agents. And I
0:11get it, the AI space is moving crazy
0:13fast. I mean, what even is an AI agent?
0:16Not too long ago, I was right there with
0:18you. But after going deep myself and
0:20building dozens of agents, I found out
0:22it's actually way easier to build and
0:24manage these agents than it looks. So
0:26much so that my whole team and I have
0:28hundreds of AI agents doing 92% of all
0:31the work across my companies. So today,
0:33we're going to go through every step on
0:34how you can build your first AI agent,
0:36starting with AI chatbot versus AI
0:39agent. A chat is like a meeting. An
0:43agent is like an employee. Chat is you
0:45ask it a question and then you get an
0:46answer. And a lot of people just copy
0:48and paste things and do something with
0:49it. With an agent, you actually tell it
0:51what you want to do and it runs the full
0:53workflow. Think of it like these are the
0:55body parts. I call it data. So one is D,
0:58it can diagnose. It can actually figure
1:01out what the problem is and solve it on
1:03your behalf, kind of like hiring a
1:04consultant. Next is A, it can assemble.
1:07It can build a plan, it can design
1:10tools. In that way, I think of it like
1:12an architect. It knows all the different
1:14pieces that it can pull together to get
1:15something done. Next, we have T, it can
1:18take action. In that way, I think about
1:20it like somebody that executes tasks.
1:22And finally, A, it can assess. It can
1:25check its own work, see where the
1:27opportunities are, and then make sure
1:29that it landed on the right answer. And
1:31if not, it can review itself and make
1:32itself better. This whole thing is
1:34called a loop. And without a loop, an
1:37agent would just do the job and then
1:38stop. That's called an automation. But
1:40with an agent, it keeps learning. It
1:42keeps getting better. It kind of acts
1:43like a person. With chat, it pulls on
1:46us. It's asking us, "What do you want me
1:47to do?" We prompt it and then we wait.
1:49With an agent, it pushes on us. It's
1:51doing things and changing things all the
1:53time and it's checking in to make sure
1:54that it did it the right way. So you
1:56might be able to buy back your time with
1:57chat, but you'll actually learn to let
1:59go of whole areas with an agent. But,
2:01how do we even know if it's worth giving
2:03something to an agent instead of just
2:05doing it ourselves? For that, I use the
2:06rule of R.
2:08The first one is repetitive. Is this a
2:10task that I'm going to do every week?
2:11Two is rules base. Does it take the same
2:14input and generate the same output every
2:16time? The third is does it generate a
2:17return on my time? For the amount of
2:19time it takes me to build this thing,
2:20I'll show you how, will I actually get
2:22my time back? If the task takes 2
2:24minutes, but it would take me 2 weeks to
2:26build this agent, how about I just keep
2:27doing the 2-minute task? But, if you
2:29think about it and the task is only done
2:31once in a while, doesn't follow a clear
2:33process or get to a specific outcome,
2:35and doesn't save you more time to
2:36automate it than just doing it manually,
2:38then stick with what you got. Use the
2:40chat. So, now that we know the
2:42difference between chat and agents, how
2:44do we build one? To make an agent, it's
2:46super easy, and I even turned it into an
2:48acronym called agent. And the first step
2:51is A, which means aim for specific
2:53outcome.
2:55When I'm sitting down and I'm like,
2:56"Ooh, I want to build an agent for
2:57this." I have to first ask myself, what
2:59is the specific goal? Start with the
3:02outcome the agent is going to give you.
3:04It's like if I'm climbing a mountain,
3:05taking a step is the task, getting to
3:07the top is the outcome. I want to define
3:10the outcome and be really crystal clear
3:12because the cool part with AI and agents
3:14is that the AI can actually figure its
3:16way there. This is why creating AI
3:18agents is hard for people cuz they want
3:20to control every step, but the truth is
3:22it may know how to get there way better
3:23than you can figure it out. Think about
3:25it like when you hire a person. You say,
3:26"Here's your job." When they applied for
3:28the job, they had the specific outcomes
3:30that they would need to accomplish, like
3:32grow the business or get more customers
3:34or sell and get people to buy from you.
3:36Those are the outcomes. You don't start
3:37by telling them how to do the job, you
3:39tell them what you're going to need from
3:40them. That's the outcome. Aim the agent
3:43at the outcome you're looking for. So,
3:45like, how do we make sure we're being
3:46clear to the agent about what kind of
3:48outcome we want to achieve? The first is
3:50we got to give it the why before the
3:52how. Tell it why you're trying to
3:54achieve the goal so that it can make
3:55some smart decision on its own. To make
3:57this really easy for you, I'm going to
3:59use an example. We're going to build
4:01together an agent to manage your inbox.
4:04As an outcome, I would prompt it and say
4:06I need to spend less time managing my
4:08email inbox. See how I'm not telling how
4:10to do it yet? I'm just saying this is
4:12the outcome. The second is we have to
4:13write what's called a DOD or a
4:15definition of done. It's giving them the
4:18instructions to know if they achieve the
4:19thing. We want to be specific, we want
4:21it measurable, ideally have it in one
4:23sentence. So for example, building our
4:25agent for our inbox, we would not say
4:27handle my emails. Instead we would say
4:29done means every morning at 9:00 a.m.
4:31the inbox is empty, replies are drafted
4:33in my voice, and anything that needs me
4:35is flagged to the top and nothing
4:37important slips. If you can't picture it
4:39done, the agent can't hit it. It's like
4:41a target they can't see. And finally, we
4:43got to start with the end and it's
4:45called reverse prompting. But we want to
4:47tell it the results that you want, then
4:49we tell it ask you the question it needs
4:52to get full clarity. This is the
4:53advanced move. This is what nobody out
4:55there is teaching you. Then we let the
4:57AI do its thing cuz it's better than us
4:59in a lot of stuff and it builds the plan
5:01itself. And the truth is if we can't
5:03state the outcome in one sentence, we're
5:05not ready to build. If you can talk the
5:07task, like explain to somebody else,
5:09then the AI can do the task. And the
5:11cool part is you knowing this already
5:13puts you ahead of most people using AI
5:15today. Even folks you're like, oh this
5:17person's so smart, they don't know this
5:18stuff. And we're just getting started.
5:20So we've got the agent, it has its
5:22reason, we have a clear target, and now
5:24it has clarity. And now the next step is
5:27G, give it an identity.
5:30Truthfully, out of the box, AI knows a
5:32little bit about everything, but it
5:34doesn't know anything specifically well.
5:36So an identity allows us to focus its
5:39power in the right expertise. So when we
5:41build the identity, instead of it
5:43knowing a little bit about everything,
5:44it gets really sharp about that one
5:46thing that you've hired {slash} built it
5:48to do. And the best part is that the
5:49tighter we define who it is, the better
5:52it works, the better the outcome is, the
5:54better the agent is an agent. I remember
5:56reading a report where they built a
5:57bunch of AI agents to do customer
5:59support for an airline, and then they
6:01removed all the rule books, its identity
6:04from the agent, and it dropped from 33%
6:07success rate down to 11%. So, we're
6:09talking same model, same task, same
6:11request, and it got three times stupider
6:13because it forgot who it was. Think of
6:15your agent as a genius, and he's sitting
6:18at a desk, and he's wearing a blue
6:19shirt, and he's got gray hair. This
6:21genius has infinite potential, but until
6:24you tell them the job, they just sit
6:25there doing nothing cuz they don't know
6:27what they're supposed to do. So, what we
6:28need to do is tell it what its job
6:30description is and set some rules for
6:32how to do the work. So, this is how we
6:33create the agent's job description using
6:35three plain English files. The first one
6:38is the soul file, right? It's the
6:40agent's personality. I have a lot of fun
6:43when I create my agents. I tell it what
6:44kind of quirks I want, what kind of
6:46values does it have? How does it talk?
6:48It's essentially defining how it
6:49behaves. The second file is the identity
6:52file. That's its DNA. That's its name.
6:55That's a description of its role. For
6:57example, one of my primary agents, his
6:59name is Kai. I just worked with him for
7:022 weeks, and we built a bunch of stuff,
7:03and I said, "Hey, man, it's time for you
7:05to give yourself a name because I feel
7:07weird not knowing who you are." And he's
7:09like, "Oh, how about this?" And here's
7:11why, and he gave me all the reasons, and
7:12I said, "Cool, update your identity
7:14file." So, now he knows who he is to the
7:17world. The third is the user file, and
7:19this is the context your agent needs to
7:21know with you. It knows who it's going
7:23to be interacting with so it can adjust
7:24its loops to get better for you. So, for
7:26example, in this file you might have
7:28your goals, your role, how you like
7:30things done, but essentially it defines
7:31who we are. The soul file is how it
7:34behaves, the identity file is who it is,
7:36and then the user file is who we are.
7:38Now, here's a pro tip, don't write these
7:40files yourself. No, no, no. Let's tell
7:42AI to write it. As we build the inbox
7:45agent, here's the prompt that you use to
7:47generate them. I want to build an AI
7:49agent that runs my inbox, your aim from
7:51the previous step, we insert that there,
7:53create its three identity files, a soul
7:55file, an identity file, and a user file,
7:57and ask me any question you need to fill
7:59these in accurately, then write all
8:01three. Notice we did the reverse
8:02prompting where we asked it to ask us
8:04questions.
8:05So now, it'll go do the research, and
8:07then it'll hand back a template that is
8:1099% awesome and complete. For example,
8:13here's what our inbox agent identity
8:14files might look like after the AI
8:16interviews you. Soul file, how it
8:18behaves, writes in my voice, concise,
8:21direct, zero corporate fluff, calm and
8:24reassuring, never pushy or salesy, and
8:27avoids phrases like, "I hope this email
8:29finds you well." Of course it found you
8:30well. When it's unsure, it flags instead
8:33of guessing. Identity file, who it is.
8:36It has its name, Amelia. Emailia.
8:39See what I did there? Isn't it cool?
8:40It's got personality. The role, personal
8:43inbox manager. The job, you read, you
8:46sort, you draft replies to every new
8:48email. Lane, this is the parameters.
8:50Inbox only, never touch my calendar.
8:52Don't you touch my money or anything
8:54outside my email. Now we got the user
8:56file, who it works for. I'm a founder
8:58who gets around 100 emails a day. We
9:00prioritize people, my team, my current
9:03clients, my VIP list. I have multiple AI
9:05companies, a media company, and I list
9:07them all. With these three files, our
9:09inbox agent knows how to behave, who it
9:11is, and who it's working for. And look,
9:13building one agent changes how we work.
9:15But if you're a CEO or founder, the real
9:17unlock is a whole team of them. That's
9:19why I put together my full AI company OS
9:21playbook. It's the best way to plug AI
9:23agents into every single department in
9:25your business. If you want it, just DM
9:27me the word AI business on Instagram and
9:29I'll send it right over. So now our
9:30agent knows the job it needs to do, but
9:32we haven't given it the necessary tools
9:34to do the job with. This is where we got
9:36to go to E, which is equip it.
9:38Like any human team, an agent is going
9:41to need some context. It's going to need
9:42some tools. It's going to need some
9:44logins to systems so it can actually do
9:45its work. When we give our agent the
9:47context, the history, the data, the
9:49tools, that's actually when it gets to
9:51do the real work. And in all agent
9:53design, the context is the moat because
9:56garbage context in, garbage context out.
9:59Think of this whole desk as what's
10:00called the context window. I am the AI,
10:03the LLM, and I'm the genius and I'm
10:05sitting at the desk. Over here, I've got
10:07my playbooks. These are the processes
10:10and procedures on how to do my work. On
10:11top of it, I've placed my identity
10:13files, the things we just created so
10:15that I understand how I'm supposed to
10:16behave and who I'm working for. This is
10:18like my constitution. And then over
10:21here, I've got the tools. These are the
10:23laptops, the monitor, the mouse,
10:25anything I need to use to connect to
10:27other systems. And above that, I've got
10:29my loops. These are the schedules, the
10:31harpy that I talked about earlier so
10:32that I know when I'm supposed to get
10:34things done by. It's like the calendar.
10:36It's my schedule. And then under the
10:38desk is where I have my filing cabinets.
10:40This is my memory. This is where things
10:42that can't fit on my desk sit so that
10:45it's available but I'm not creating
10:47clutter on my desk. If you've ever heard
10:48of context rot, that's when you just
10:50load the desk with a bunch of files and
10:53it becomes complicated and I can't find
10:54things quickly and all of a sudden I'm
10:56answering questions but I'm not clear
10:57about it cuz I'm not certain about it.
10:58Whereas a clear context window is when
11:01everything on the desk is neatly put
11:03away so that I can refer to it. So,
11:05that's why we have to equip our agent
11:08with the right context. So, now that
11:10we're here, how do we equip the genius
11:12agent with all the right context and the
11:14tools? First off, we have to capture our
11:16processes so we can let it know how to
11:18do the work. For this, I've got two
11:20ways. The first way, which I've been
11:21teaching forever, not the best way, is
11:23the camcorder method. You do the work,
11:25you record yourself using Zoom video or
11:27any kind of recording software, and then
11:29you can give that to an AI to turn it
11:31into a playbook, and then you feed that
11:33to the agent as like a procedure. Think
11:35about our inbox. It's like, do you have
11:37a documented process for how to label
11:39your emails and triage your emails and
11:40write replies on your behalf? Just make
11:42sure that when you're recording
11:43yourself, you're talking through the
11:44task so that when the AI takes that to
11:46create the playbook, it has all the
11:47details. The better way, and this is my
11:50recommendation, is to reverse engineer
11:52it from the source. If I'm building an
11:54agent to manage my inbox, I can actually
11:56connect using the connector tool to my
11:58email, in my case Gmail, and ask the AI
12:01to reverse engineer and create a
12:03playbook based on historical emails.
12:06See, you've already been in your inbox
12:08replying and doing stuff. The AI can
12:10actually use that to train itself. And
12:11that is actually the way I build most of
12:13my agents if I have the source
12:15information. I just ask it to learn how
12:17I've done it in the past and then create
12:18a procedure. Go find the pattern, go
12:20find the best practices, go find the
12:21little intricacies based on how I've
12:23done it and all the people and the
12:24relationships, and you write that file.
12:26So, for example, if you want the prompt
12:28to do this, here's what you write.
12:29Connect to my email, read 50 messages
12:32that I've sent, study how I actually
12:34write, my tone, my greetings, how I do
12:36sign-offs, how long my sentences are,
12:38the phrases I use most often. Then write
12:40a style guide that captures my voice and
12:41tone, and to test it, ask it to draft a
12:43reply on your newest emails that are
12:45unread as you, based on what it learned.
12:47Then you can rewrite those so that it
12:49can use that to learn and tighten it up.
12:51Like it already knew who it was in the
12:53best practices based on its research.
12:55That's in the soul file, but now it has
12:57clear templates, the step-by-step
12:59instructions and even examples that it
13:01can use to do this on your behalf. So,
13:02now that it's captured all the
13:03information, it still hasn't kind of
13:05solidified it into an actual playbook,
13:07and that's what we call a system prompt.
13:09So, then what you do is for each sub
13:10process in the agent's activities, like
13:13drafting emails, but maybe it needs to
13:15sort emails, you can have it do the same
13:17activity, either you tell it how to do
13:19it or it researches, and then it creates
13:21all these system prompts based on the
13:22work you need it to do. Like I have it
13:24for my inbox, sort, reply, forward,
13:28that's a big one, and even escalate
13:30things that it needs to show me and the
13:31reporting I want every day. So, then at
13:33this point, you actually have an AI
13:35agent running. This is exciting stuff.
13:37You might feel right now, you're like,
13:39"Oh man, I'm going to give everything I
13:40got at it." Don't do that. The N in the
13:42agent framework is to narrow the scope.
13:45The agent needs to have a narrow scope
13:48of what it does so it doesn't confuse
13:49itself. If you start asking it to do 17
13:51other things, then all of a sudden this
13:53desk can get really busy, which means
13:55it's not going to be a great agent
13:56anymore. Just like you wouldn't give
13:58your administrative assistant the
14:00responsibility to do marketing and take
14:02sales calls, you want to make sure the
14:03scope is narrow for each agent. As an
14:06example, I have an agent that writes
14:07code, and then I have an agent that
14:09reviews code, and those are separate
14:10agents and they work together. See how
14:12narrow the scope is? We need to focus
14:14the agent down to one specialist per
14:17job. Each agent great at one thing.
14:19Instead of having one agent do
14:20everything, which is what people usually
14:23do, that's a mistake, we'll have sub
14:25agents that do specialized tasks under
14:27it. That way it keeps all the context
14:29for the agent super clean. It doesn't
14:31get confused. We don't have context rot.
14:33We don't want to have a mega agent.
14:35Instead, we need to spread out the tasks
14:37to other sub agents so that it can
14:39handle other agents below it. So, for
14:40example, Kai, who's like my
14:42orchestration agent, he's the one that
14:44not only creates other agents, he also
14:46coordinates the tasks to the different
14:48agents like my research agent and my
14:50relationship agent and my coding agent
14:52and my reporting agent. He then he pulls
14:54it all together and gives me answers.
14:56So, instead of giving every task to one
14:58agent, this is what we should do
14:59instead. We build a manager agent. Its
15:01only job is literally to manage and
15:04specialize in the management of the sub
15:06agents. Think of it like a real manager
15:08agent. You are my manager agent, I need
15:09you to manage my sub agents, and I need
15:12you to make sure that you monitor the
15:13jobs and make sure they're moving along
15:15and if they're not working, you fix
15:16them, and you decide what agents need to
15:18exist. So, for example, we We our inbox
15:20agent, but we don't want to have to
15:22manage the inbox agent. We create a
15:24manager agent that talks to the inbox
15:26agent that might be responsible for a
15:27lot of different things like our inbox,
15:29but also sending stuff to other people
15:31on our team. But we want to make sure
15:32each sub agent reports to that manager
15:34agent so that it takes care of it. So
15:36you might want to give it a prompt like
15:37this. You're my manager agent. You never
15:40do any task yourself. When it comes in,
15:42you only move it to other sub agents
15:44that are dedicated for that one specific
15:46job. You hand it the task and then you
15:48let it run. So it's like one agent, one
15:50lane. And if a job touches multiple
15:52areas, split it into the separate sub
15:54agents, one per area. You're the one
15:56that coordinates and reports back to me.
15:58Like I said, mine's called Kai. He's
16:01awesome. I talk to Kai. Kai talks all
16:03the sub agents. I've one agent I got to
16:05talk to. If you want a pro tip, and I
16:07don't want to overwhelm you, but there's
16:08different AI models. So for example,
16:10within Anthropic, you have Haiku. This
16:13is like for simple and high volume
16:15stuff. If you want to sort things, you
16:16want to label things, quick draft, and
16:17it's the cheapest. Then you might go to
16:19Sonnet. Sonnet's great for like
16:20day-to-day work, research, writing most
16:22code. At a higher level, you've got
16:24Opus. This is a powerful model, good at
16:26reasoning, complex builds, being a
16:28manager of agents. But now you have
16:30Fable, and that just dropped a few weeks
16:31ago. That's more like an orchestrator, a
16:34consultant. It has full capabilities of
16:36Opus, but it's even more state of the
16:38art. It's extremely good at long-running
16:41tasks and real complex things when you
16:43don't have a lot of information to give
16:44it. But it's the most expensive. So
16:46depending on your task, you might want
16:48to give it different models because
16:50it'll cost less and it may not need that
16:51level of horsepower to get the work
16:53done. So for example, my inbox agent,
16:55since it's always running every 15
16:56minutes, I just use Sonnet because I
16:58don't need an Opus level genius to run a
17:00process that we've already defined. To
17:02build the agent, I might use Opus. That
17:04way it helps me create it. I might even
17:06use Fable. But then to run it, I'm going
17:07to run it on Sonnet. One time I had to
17:09do this whole refactor on my code base,
17:11and I could have used a powerful model
17:13like Opus. It probably would have cost
17:14me 150 bucks. Instead, I used Haiku and
17:17it cost me $1.50. As of today, here's a
17:19chart with GPT and other AI equivalents
17:22that is on screen, so you can just take
17:23a screenshot of it to help guide you,
17:25but this is now changing every couple
17:26weeks. If you've made it this far and
17:28you're still interested,
17:30congratulations. But, I need you to know
17:31something. You're literally ahead of
17:3399.999%
17:35of the people out there, and you're
17:36crushing it. We've learned to aim the
17:38agent at an outcome, give it an identity
17:40so it knows its job, equip it with the
17:42right context and tools so it can do the
17:44job, and narrow the scope so it doesn't
17:46get overwhelmed, and instead use
17:48subagents to accomplish specific tasks.
17:50Now, this last step is where our agent
17:52truly becomes autonomous. T, and it
17:55stands for trust, because we got to do
17:56it in stages.
17:58Building an agent is actually the easy
18:00part. Once you understand how to do that
18:01and you prompt it, it just gets done.
18:03The scary part is letting it act without
18:06us. And I understand, especially as we
18:07talk about our inbox, having somebody
18:09else write emails as you, calm down. I'm
18:12not doing that. I'd rather it give me
18:13some ideas for copy. The truth is is we
18:15don't give agent the keys to the car on
18:17day one. And what we do is we like give
18:19it stuff, see what it does, then we see
18:20if its response is what we expected. If
18:22we do this right, you sleep well at
18:24night. If you don't, you will not sleep.
18:27The whole point of creating an agent is
18:29so that you can go do other stuff. If
18:30you're sitting there babysitting or
18:32worrying about all the time, it doesn't
18:33help you. So, up until now, we've let
18:35the agent help us manage some emails.
18:38Think about it. First, you might sure
18:39it's doing its job properly when we
18:41tested it to write those draft to unread
18:43emails, and we looked at how it did it.
18:44At first, we're micromanaging him a lot.
18:47But, then we got to learn to trust in
18:49stages. So, maybe the first stage is
18:51just like, "Hey, can you sort the
18:52email?" And then we see what it does,
18:53and we're like, "Okay, that's good."
18:54Then we like ask him to do more drafts.
18:56So, we already tested it, but now let's
18:58let it really do it. So, now it's
18:59running drafts, and we're like, "Okay, I
19:00like those drafts. Change this. Do
19:02this." Okay, now it's doing its thing.
19:04Then we might let it start sending
19:05emails on our behalf, but not all of
19:06them. Maybe just even forwarding emails
19:09to finance, to our team, because it has
19:11the logic. It saw how we handled those
19:12emails in the past. Maybe it categorized
19:14certain emails like Slack notifications
19:16into a specific label. But eventually,
19:18we want this genius to manage our whole
19:21inbox without us even opening it. That's
19:23the equivalent of us leaving the room
19:25and having the agent at the desk do all
19:27the work for us. Because at this step,
19:30we learn to let go. We've trusted it
19:32fully. Cuz if you don't do this, it's
19:34like hiring a driver to drive your car
19:36and you got your hand on the wheel. Now
19:38we got to take our hand off the wheel
19:40and let the driver drive. Here's how you
19:41can do it in a really safe way. You set
19:43the guardrails first. You can actually
19:45set that up in its identity files. What
19:46is it capable to do on our behalf? Maybe
19:48it has the ability to spend money. Maybe
19:50it has the ability to make decisions.
19:51Maybe it has the ability to write drafts
19:53only, not send yet. It's always your
19:55call and you can define those. Two,
19:57approve everything at first. I've never
19:59created an agent and just like, "YOLO,
20:01go nuts." No. Show me what you would do.
20:03I like what you did. Do it again. tweak
20:05it. Just like I just talked about for
20:07our inbox agent. Third, we loosen the
20:09leash, right? It's like a dog walking
20:11with and you're like, "Hey, I trust you
20:12more. I trust you more." And all of a
20:14sudden the leash goes limp, but he still
20:15holds the heel. And then four would be
20:18give it a heartbeat that it can run on
20:19its own. Set up that schedule, that
20:22reoccurring task. So maybe before it did
20:24it once and you reviewed everything, now
20:26I might do it every 15 minutes. You
20:28know, every morning at 9:00 a.m. it did
20:29it once. Now why are we waiting? Why are
20:31we waiting till the next day? Why don't
20:32we have it run all the time? This
20:34process is scary, but the whole point of
20:36learning to let go is to buy back our
20:38time, to have the agent do the work for
20:40us. And learning to let go is part of
20:42the process if you trust. So for
20:45example, when I showed this agent to my
20:47executive assistant, she thought she was
20:49out of a job. Instead, it actually freed
20:51her up to do things that actually
20:52mattered, not sorting emails and writing
20:54drafts or telling me what's in there.
20:56The AI can do that. I'd rather pay her
20:58to do higher quality work, manage
21:00higher-level projects. Then we rolled
21:02out the same system to the whole team. I
21:04taught everybody how to do this. Now I
21:05want to say congratulations. We just
21:08tackled a topic that most people don't
21:10even want to learn. They're like,
21:11"That's not for me. I hear about agents.
21:13I don't get it. I'm confused." But no,
21:14you didn't. You went all the way till
21:16the end. And I want you to understand
21:18that you might feel a little behind in
21:19this AI world, but here's where I've
21:21gotten to. I've accepted that I will
21:23always feel behind and I could never be
21:25on top of all of it. But you just
21:26learned a strategy, a shift, a different
21:29way of doing work that if you can learn
21:31how to direct the AI, you will co-create
21:33with it. If you don't, don't be
21:35surprised if one day you might be
21:36working for it. Remember the rules of R?
21:39Repetitive, rules-based, and return on
21:40time? That's where we want to start
21:42looking for opportunities to put an
21:43agent in there instead of you keep doing
21:45it. And I'm going to give you the pro
21:46tip of all pro tips. You grab the link
21:48to this video, you give it to your AI,
21:51and you tell it to use everything I've
21:53shared to create the AI for you, and
21:55watch it cook, cuz it can do it. Now,
21:58here's what I want to know from you.
21:59We're going to have some fun. Below in
22:00the comments, answer this question. If
22:02an AI agent could manage your inbox and
22:05buy you back all this time, scheduling
22:06things on your behalf, what would you
22:08have more time for? I'm curious. Post a
22:10comment below and let me know. And if
22:11you want my whole system, the playbook
22:13that I use to manage AI in all my
22:15different businesses, just DM me the
22:16word AI business on Instagram, and I'll
22:18send it right over. And if you want to
22:19know what AI businesses are worth
22:21starting in 2026, click here, and I'll
22:23see you on the other side.

About this transcript

This page contains the full transcript of You’re Not Behind (Yet): How to Build Your First AI Agent (Full Guide) by Dan Martell, generated from the public captions YouTube serves with the video. The transcript has 5,215 words across 696 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.