AI Agents EXPLAINED in 14 minutes and TOOLS for building one — Silicon Valley Girl Podcast

Marina Mogilko: AI agents, AI agent. My team and I built a real AI agent. Yes, we're trendy. Yes, we're doing this. And it actually automated a process that used to be very manual. The process is called building like a content factory where we have one video, but we want to make a lot of different content pieces from it. And this AI agent is already running and it's working instead of us. It feels like we just added another person to the team. And this person costs maybe like 40 bucks a month that we spend on all the automation tools. In this video, I'll show you how regular AI tools turn into agents and fully automate content from video to publishing. I'll show you our working system and I want you to remember the tools that we're using because they're very universal. Some automation tools can not only work for content, they can work for anything. And I know that my team at Lingu uses these same tools to automate other things in the company. But basically what's happening, we're in the era of AI agents. It's not just like talking to ChatGPT and getting an answer. It's actually asking an AI agent to repeat a process. And the best thing is you don't need to be a developer to understand how it works in practice.

So there are three different levels. Level one, regular neural networks, ChatGPT, Gemini, Claude. Then you have AI workflows where you build the logic yourself. What happens when, and why, and AI does it for you. And then there is this cool thing called AI agents where AI thinks, decides, and acts on its own. And most importantly, I'll show you how our AI agent takes a video, cuts it, writes titles, and publishes shorts without human involvement. So, let's start with level one. It is where you ask a question and an LLM gives you an answer. Simple principle, input, output. For example, if I ask my ChatGPT, give me a list of the top five AI tools for YouTube, I get high quality, up-to-date answer. But once you ask, what videos did I publish on YouTube over the last three days? The model cannot answer. Why? Because it doesn't know who you are. It has no access to your YouTube channel, email calendar, unless it's Gemini. Gemini can answer questions about your email, but then you ask Gemini about the weather on a particular day and it's like it freezes. LLM is an isolated system that doesn't see your personal data and that's actually a privacy bonus. But most importantly, these models are passive. They wait for you to give a command and they don't do anything on their own. They can't act. They don't make decisions. They don't build action chains. It's still just a tool. Smart, useful, but not an agent.

Then we move to level two, AI workflow. At this second level, you're no longer just chatting with a neural network. You start automating. This is where magic starts happening. You build a logical chain of actions yourself. If A, then do B and C. This is called an AI workflow. So here's our real life example. We use apps every day. Weather, Google calendar, notes, email. What if you tell the AI if I ask about my personal events, check my Google calendar first and only then answer. Now, if I ask when is the meeting with Reed Hoffman, the AI checks the calendar and gives the correct answer. But if I immediately follow up with what will the weather be like that day, it fails again because you did not add the step check the weather. It simply doesn't know it has to check the weather. And that's how a workflow operates. You build the route, the AI follows it. No flexibility, no adaptation, but it still does things by itself. And this is called control logic. As long as you decide what to do and when, it's not an agent. It's just a manually built process, which is again cool. But let's talk about AI agents.

So the entire system that we built, we used a platform called Nathan, an automation tool similar to Make or Zapier but even more flexible. It allows you to visually create a chain of actions between different services without writing a single line of code. Here is how our workflow works. All links to uploaded videos are stored in Google Sheets. This is our starting point where the workflow pulls the new tasks. The Nathan scenario connects Google Sheets and retrieves the most recent uploaded long-form video link. This video is automatically sent to Clap AI, a neural network powered tool that converts long videos into vertical shorts. It detects key moments, emotional peaks, and generates clips in just a few minutes. But we don't stop here. Then we ask ChatGPT to analyze the content of each short and generate viral titles and descriptions optimized for YouTube shorts. We feel like it does it better than the tool itself. We use a custom prompt to ensure the titles are catchy, concise, and designed for maximum reach. And the final step is automated upload of shorts to YouTube. The workflow publishes up to 10 videos per day, saving us a ton of time. It uploads files, adds the generated titles and descriptions, and posts the content to the channel. This basically gives us full automation without any human involvement. No editors, no managers, no manual work. Every day, our AI powered system publishes content that drives views. That means free traffic without spending on ads. This approach allows you to scale yourself, even if you're working solo. And that's a real AI agent.

Now, let me explain the key difference between an AI agent and a workflow that we just talked about. So, as I mentioned, a workflow is a set of steps that follow a strict script. But a true AI agent is more than just execution. It can actually think, act and adapt. So we have three capabilities. First of all, reasoning. It thinks. The AI agent chooses how to solve the task on its own. For example, instead of just grabbing the latest file, it checks which video is most likely to go viral. It acts. It operates tools. It doesn't wait for a command. It connects to a necessary service, looks for information, and launches actions on its own. And it iterates. It refines the result. If the first solution doesn't work, it tries again. For example, it generates three versions of a title and chooses the one with the highest expected CTR, or it sends the result to another LLM to get some feedback and it improves based on that feedback. The logic is called ReAct. Reason plus act is not just a task sequence. It's a process where the neural network becomes the brain, not just a tool.

In a regular workflow, if you don't like the result, you go back and manually change the prompt. For example, if a post that your workflow creates turns out boring, you rewrite the request and rerun the workflow. Now, imagine the AI agent does all of that on its own. In our scenario, it might look like this. GPT generates a title for a short. It then sends that title to another LLM, for example, with the role of a YouTube editor, like trainer ChatGPT or whatever. The second LLM analyzes it and says, "Huh, this title is too generic. The CTR will be low." The agent returns to generation and creates a new version. It repeats the cycle until all conditions are met—short specific with a hook. And the most important part, the human is no longer involved. And this is what makes the system agentic. The AI sets the goal, evaluates intermediate results, and independently decides what to do next.

I did not build it by myself. I am less technical than that. We actually hired someone. We call them AI interns. So funny. I was talking to my friend a couple months ago when summer was just starting and she told me, "This summer I'm working with an intern who did three years of computer science or something. They're on their summer break and they're going to automate what my assistant does. She has a legal firm." And I was like, "This is genius. We don't necessarily need a professional developer, but would totally work with a couple of interns." And so that's what we did. We hired a couple interns and we talked to them about the processes that are very manual right now. And I want my team to focus on something creative like who is our next guest? What are we going to make a video about? What are our big goals? I don't want them to just manually publish things. We still do a lot of things manually, but it's a process.

I talked to someone from the Diary of a SEO team and they make so many pieces of content every single day for many, many different channels. I'm like, I want this system, but I can't afford it if I just hire people. But the solution was an AI agent. So, we built one and it's improving. It's exciting. And as an entrepreneur, when I see that with the same team, our results are not 10x right now, but like 5x, my natural response to that is like, let's 100x. How do we do that? Hire more people. Perfect. So, we're still hiring. Even with all the AI agents, we just need more people because I want to grow, right? It's funny how this mindset works—oh, yes, we're replacing a person, but we actually need a couple more people because we want this to scale.

I also wanted to show you a fascinating example of an AI agent that was created by OpenAI because they've recently launched this viral ChatGPT agent—I think you've heard of it—and it's no longer just a smart chatbot, it's actually an agent capable of completing goal-oriented tasks independently. Like visiting websites, gathering and analyzing information, filling out forms. Finally, someone is filling out forms! As a person with a passport that required a lot of visas, I'm like finally guys, 30% of my business was filling out forms for clients to get a US visa or a UK visa. Finally, it works with Google calendar. It creates spreadsheets, presentations, and even purchases tickets. I want an AI agent that knows how to utilize miles and fly business class. Anyways, it uses a virtual computer with a browser, terminal, API access, and memory. It chooses the order of steps on its own, adjusts the path if something goes wrong, and it can repeat the cycle until it reaches the desired outcome.

For example, find the most viral YouTube shorts in the AI niche this week. Analyze their titles, captions, thumbnails, and generate a content plan with hooks, titles, and video ideas for my channel. And the agent goes to YouTube, collects data on top performing videos, analyzes their structure and metadata, creates an Excel content plan with actionable insights, and sends the result back to you. And all that without any additional prompts at each step. That's what reasoning plus acting plus iteration looks like. And that's exactly the kind of logic that separates an agent from regular automation.

Right now, the workflow that we built is very predictable. It executes every step precisely. Google Sheets, Clap AI, GPT, YouTube. But the logic is still set manually by us. To become a true agent, the system also needs to learn how to choose videos to process on its own. So instead of just picking a link to the video, ideally I'd like it to go to my whole YouTube channel, analyze metadata, audience behavior, and trending topics, and then generate shorts, test multiple title variations, tell me like, Marina, you need to make a video about this, this, and that. Analyze engagement, and then publish the best one.

You know, I was talking to a guy who built a system where you can test ads in an AI world, like you create an ad and then you launch it to this AI audience and you can kind of predict the outcome. This is where we're going. I feel like we're going to create AI versions of ourselves to test ideas so that every video that we publish on YouTube goes viral instantaneously because we tested it on the AI version of our audience. And ideally the agent also learns, like okay, this video we thought it was going to go viral, it did not. Why didn't it go viral? Let's adjust the workflow. Let's adjust the process.

So the goal of this video is to inspire you to automate something in your workflow and start simple. Like with hiring, you start with hiring someone who just does personal assistant type of work with this. Try to automate one repetitive task you always do by hand. I don't like typing anymore. I'm talking to my computer and I'm talking a lot and I'm becoming more productive. Ask yourself, what can I automate in my workflow? Maybe you start with something simple and then move on to building your own AI agent.