AI Agents, Clearly Explained

[00:03] AI. AI. AI. AI. AI.

[00:07] AI. You know, more agentic. Agentic

[00:10] capabilities. An AI agent. Agents.

[00:12] Agentic workflows. Agents. Agents.

[00:15] Agent. Agent. Agent. Agent. Agentic.

[00:19] All right. Most explanations of AI

[00:20] agents is either too technical or too

[00:23] basic. This video is meant for people

[00:26] like myself. You have zero technical

[00:28] background, but you use AI tools

[00:30] regularly and you want to learn just

[00:33] enough about AI agents to see how it

[00:36] affects you. In this video, we'll follow

[00:38] a simple one, two, three learning path

[00:41] by building on concepts you already

[00:43] understand like chatbt and then moving

[00:46] on to AI workflows and then finally AI

[00:49] agents. All the while using examples you

[00:52] will actually encounter in real life.

[00:55] And believe me when I tell you those

[00:56] intimidating terms you see everywhere

[00:58] like rag, rag, or react, they're a lot

[01:02] simpler than you think. Let's get

[01:04] started. Kicking things off at level

[01:05] one, large language models. Popular AI

[01:08] chatbots like CHBT, Google Gemini, and

[01:10] Claude are applications built on top of

[01:14] large language models, LLMs, and they're

[01:17] fantastic at generating and editing

[01:19] text. Here's a simple visualization.

[01:21] You, the human, provides an input and

[01:24] the LLM produces an output based on its

[01:27] training data. For example, if I were to

[01:29] ask Chachi BT to draft an email

[01:31] requesting a coffee chat, my prompt is

[01:33] the input and the resulting email that's

[01:36] way more polite than I would ever be in

[01:37] real life is the output. So far so good,

[01:40] right? Simple stuff. But what if I asked

[01:43] Chachi BT when my next coffee chat is?

[01:47] Even without seeing the response, both

[01:49] you and I know Chachi PT is gonna fail

[01:52] because it doesn't know that

[01:53] information. It doesn't have access to

[01:56] my calendar. This highlights two key

[01:58] traits of large language models. First,

[02:00] despite being trained on vast amounts of

[02:02] data, they have limited knowledge of

[02:04] proprietary information like our

[02:07] personal information or internal company

[02:09] data. Second, LLMs are passive. They

[02:12] wait for our prompt and then respond.

[02:14] Right? Keep these two traits in mind

[02:17] moving forward. Moving to level two, AI

[02:19] workflows. Let's build on our example.

[02:21] What if I, a human, told the LM, "Every

[02:25] time I ask about a personal event,

[02:26] perform a search query and fetch data

[02:29] from my Google calendar before providing

[02:31] a response." With this logic

[02:33] implemented, the next time I ask, "When

[02:35] is my coffee chat with Elon Husky?" I'll

[02:38] get the correct answer because the LLM

[02:40] will now first go into my Google

[02:42] calendar to find that information. But

[02:45] here's where it gets tricky. What if my

[02:48] next follow-up question is, "What will

[02:50] the weather be like that day?" The LM

[02:53] will now fail at answering the query

[02:55] because the path we told the LM to

[02:57] follow is to always search my Google

[03:00] calendar, which does not have

[03:02] information about the weather. This is a

[03:04] fundamental trait of AI workflows. They

[03:07] can only follow predefined paths set by

[03:10] humans. And if you want to get

[03:12] technical, this path is also called the

[03:15] control logic. Pushing my example

[03:17] further, what if I added more steps into

[03:20] the workflow by allowing the LM to

[03:22] access the weather via an API and then

[03:24] just for fun use a text to audio model

[03:26] to speak the answer. The weather

[03:28] forecast for seeing Elon Husky is sunny

[03:31] with a chance of being a good boy.

[03:33] Here's the thing. No matter how many

[03:35] steps we add, this is still just an AI

[03:39] workflow. Even if there were hundreds or

[03:41] thousands of steps, if a human is the

[03:44] decision maker, there is no AI agent

[03:47] involvement. Pro tip: retrieval

[03:49] augmented generation or rag is a fancy

[03:52] term that's thrown around a lot. In

[03:54] simple terms, rag is a process that

[03:56] helps AI models look things up before

[03:58] they answer, like accessing my calendar

[04:00] or the weather service. Essentially, Rag

[04:03] is just a type of AI workflow. By the

[04:06] way, I have a free AI toolkit that cuts

[04:07] through the noise and helps you master

[04:09] essential AI tools and workflows. I'll

[04:10] leave a link to that down below. Here's

[04:12] a real world example. Following Helena

[04:14] Louu's amazing tutorial, I created a

[04:17] simple AI workflow using make.com. Here

[04:19] you can see that first I'm using Google

[04:21] Sheets to do something. Specifically,

[04:23] I'm compiling links to news articles in

[04:25] a Google sheet. And this is that Google

[04:28] sheet. Second, I'm using Perplexity to

[04:31] summarize those news articles. Then

[04:34] using Claude and using a prompt that I

[04:36] wrote, I'm asking Claude to draft a

[04:38] LinkedIn and Instagram post. Finally, I

[04:42] can schedule this to run automatically

[04:44] every day at 8 a.m. As you can see, this

[04:46] is an AI workflow because it follows a

[04:49] predefined path set by me. Step one, you

[04:52] do this. Step two, you do this. Step

[04:55] three, you do this. And finally,

[04:57] remember to run daily at 8 am. One last

[04:59] thing, if I test this workflow and I

[05:02] don't like the final output of the

[05:05] LinkedIn post, for example, as you can

[05:08] see right here, uh, it's not funny

[05:10] enough and I'm naturally hilarious,

[05:11] right? I'd have to manually go back and

[05:16] rewrite the prompt for Claude. Okay? And

[05:20] this trial and error iteration is

[05:23] currently being done by me, a human. So

[05:25] keep that in mind moving forward. All

[05:27] right, level three, AI agents.

[05:29] Continuing the make.com example, let's

[05:31] break down what I've been doing so far

[05:33] as the human decision maker. With the

[05:36] goal of creating social media posts

[05:37] based off of news articles, I need to do

[05:39] two things. First, reason or think about

[05:43] the best approach. I need to first

[05:44] compile the news articles, then

[05:46] summarize them, then write the final

[05:48] posts. Second, take action using tools.

[05:51] I need to find and link to those news

[05:53] articles in Google Sheets. Use

[05:55] Perplexity for real-time summarization

[05:58] and then claw for copyrightiting. So,

[06:00] and this is the most important sentence

[06:01] in this entire video. The one massive

[06:04] change that has to happen in order for

[06:06] this AI workflow to become an AI agent

[06:09] is for me, the human decision maker, to

[06:13] be replaced by an LLM. In other words,

[06:16] the AI agent must reason. What's the

[06:19] most efficient way to compile these news

[06:20] articles? Should I copy and paste each

[06:22] article into a word document? No, it's

[06:24] probably easier to compile links to

[06:26] those articles and then use another tool

[06:28] to fetch the data. Yes, that makes more

[06:30] sense. The AI agent must act, aka do

[06:34] things via tools. Should I use Microsoft

[06:37] Word to compile links? No. Inserting

[06:39] links directly into rows is way more

[06:41] efficient. What about Excel? M. So the

[06:44] user has already connected their Google

[06:45] account with make.com. So Google Sheets

[06:47] is a better option. Pro tip. Because of

[06:49] this, the most common configuration for

[06:51] AI agents is the react framework. All AI

[06:55] agents must reason and act. So

[06:59] react. Sounds simple once we break it

[07:01] down, right? A third key trait of AI

[07:03] agents is their ability to iterate.

[07:06] Remember when I had to manually rewrite

[07:08] the prompt to make the LinkedIn post

[07:10] funnier? I, the human, probably need to

[07:13] repeat this iterative process a few

[07:15] times to get something I'm happy with,

[07:17] right? An AI agent will be able to do

[07:19] the same thing autonomously. In our

[07:22] example, the AI agent would autonomously

[07:25] add in another LM to critique its own

[07:28] output. Okay, I've drafted V1 of a

[07:30] LinkedIn post. How do I make sure it's

[07:32] good? Oh, I know. I'll add another step

[07:34] where an LM will critique the post based

[07:36] on LinkedIn best practices. And let's

[07:38] repeat this until the best practices

[07:40] criteria are all met. And after a few

[07:42] cycles of that, we have the final

[07:45] output. That was a hypothetical example.

[07:47] So let's move on to a real world AI

[07:50] agent example. Andrew is a preeeminent

[07:53] figure in AI and he created this demo

[07:55] website that illustrates how an AI agent

[07:58] works. I'll link the full video down

[08:00] below, but when I search for a keyword

[08:02] like skier, enter the AI vision agent in

[08:07] the background is first reasoning what a

[08:10] skier looks like. A person on skis going

[08:12] really fast in snow, for example, right?

[08:14] I'm not sure. And then it's acting by

[08:18] looking at clips in video footage,

[08:22] trying to identify what it thinks a

[08:24] skier is, indexing that clip, and then

[08:29] returning that clip to us. Although this

[08:32] might not feel impressive, remember that

[08:34] an AI agent did all that instead of a

[08:36] human reviewing the footage beforehand,

[08:39] manually identifying the skier, and

[08:42] adding tags like skier, mountain, ski,

[08:45] snow. The programming is obviously a lot

[08:47] more technical and complicated than what

[08:49] we see in the front end, but that's the

[08:51] point of this demo, right? The average

[08:53] user like myself wants a simple app that

[08:56] just works without me having to

[08:58] understand what's going on in the back

[09:00] end. Speaking of examples, I'm also

[09:02] building my very own basic AI agent

[09:05] using Nan. So, let me know in the

[09:07] comments what type of AI agent you'd

[09:08] like me to make a tutorial on next. To

[09:11] wrap up, here's a simplified

[09:12] visualization of the three levels we

[09:14] covered today. Level one, we provide an

[09:17] input and the LM responds with an

[09:19] output. Easy. Level two, for AI

[09:22] workflows, we provide an input and tell

[09:24] the LM to follow a predefined path that

[09:27] may involve in retrieving information

[09:29] from external tools. The key trait here

[09:31] is that the human programs a path for LM

[09:34] to follow. Level three, the AI agent

[09:37] receives a goal and the LM performs

[09:39] reasoning to determine how best to

[09:41] achieve the goal, takes action using

[09:44] tools to produce an interim result,

[09:46] observes that interim result, and

[09:48] decides whether iterations are required,

[09:51] and produces a final output that

[09:53] achieves the initial goal. The key trait

[09:56] here is that the LLM is a decision maker

[09:58] in the workflow. If you found this

[10:00] helpful, you might want to learn how to

[10:02] build a prompts database in Notion. See

[10:04] you on the next video. In the

[10:05] meantime, have a great one.