Like many of you juggling busy schedules, I often find myself needing to grasp new concepts quickly. While videos can be great, sometimes I’m just more of a “reading kind of guy.” I like to digest information at my own pace, maybe reread a section, or easily refer back to it later .
Recently, I stumbled upon this excellent YouTube video by the channel “Jeff Su” that breaks down AI agents in a really clear, step-by-step way. It uses concepts many of us already understand, like ChatGPT, and builds up from there.
I thought it would be useful to embed the video right here at the start. Give it a watch if you have a moment!
AI Agents, Clearly Explained :
Watched it? Great! Didn’t have time? No worries. For fellow readers, or if you just want my perspective layered on top, let’s walk through the key ideas presented in the video together. I found its “Level 1, 2, 3” approach really helpful.
Level 1: The Foundation – Large Language Models (LLMs)
The video starts with something most of us are familiar with: Large Language Models or LLMs. These are the engines behind popular tools like ChatGPT, Google Gemini, and Claude – tools I personally use almost daily for everything from drafting emails and brainstorming content ideas for this blog to even getting help refining code snippets for automation tasks.
The core idea, as the video visualizes, is simple: You (Human) provide Input -> LLM processes -> LLM produces Output (based on its training data).
Think about asking ChatGPT to write a polite follow-up email. Your request is the input, and the generated email is the output.
But, as the video rightly points out, standard LLMs have limitations. They generally can’t access your personal, real-time information (like your calendar) or confidential company data. They also operate passively – they wait for your command and then respond. This is crucial to remember as we move to the next level.
Level 2: Taking a Step Forward – AI Workflows
This is where things start getting more interesting and closer to the automation I’ve worked with in my digital marketing career, particularly during my time managing campaign systems at SMARTFREN.
An AI workflow, as explained in the video, is essentially telling an LLM to follow a predefined path that might involve interacting with external tools or data sources.
The video uses the example: “When I ask about a personal event, first check my Google Calendar.” Now, the LLM isn’t just responding based on its training data; it’s following a specific instruction to fetch external information before generating the output.
This might involve multiple steps:
- Get data from Google Sheets.
- Summarize it using Perplexity AI.
- Draft a social media post using Claude with a specific prompt I wrote.
- Schedule it to run daily.
(Suggested Image: A simple flowchart diagram illustrating a sequence: Step 1 -> Step 2 -> Step 3 -> Output.)
This sounds powerful, and it is! It’s the basis for a lot of marketing automation. We set up triggers and actions – if a user clicks this link, send them that email sequence; if a campaign reaches this threshold, update that report.
The video mentions Retrieval-Augmented Generation (RAG) here. Don’t let the fancy term scare you. As the video simplifies, RAG is basically a common type of AI workflow where the AI is instructed to “look things up” (retrieve information from a specific source, like my calendar or a company knowledge base) before answering (generating a response).
However, the key defining trait of an AI workflow is that the human is still the decision-maker. I defined the path. I wrote the prompts. If the output isn’t right (like the LinkedIn post in the video example not being funny enough), I have to go back and manually tweak the instructions or prompts. This trial-and-error iteration is human-led.
Level 3: The Leap – AI Agents
This is the main event! What makes an AI Agent different from an AI Workflow?
The video nails the core distinction with what it calls “the most important sentence”: The massive change is replacing the human decision-maker with an LLM.
Instead of just following my predefined path, the AI Agent is given a goal and then needs to:
- Reason: Figure out the best way to achieve the goal. What steps are needed? What’s the most efficient approach? (e.g., “Should I copy-paste articles, or compile links and use a tool? The second option seems better.”)
- Act: Use available tools to execute those steps. (e.g., “Okay, I need to compile links. Google Sheets is connected and efficient for this.”)
- Observe & Iterate: Look at the result of its actions. Is it good enough? Does it meet the goal? If not, figure out how to improve it and try again autonomously. (e.g., “This draft isn’t quite right based on best practices. Let me critique it using another model and refine it.”)
The video mentions the ReAct framework (Reason + Act) being common for agents. Once broken down like this, it makes perfect sense, right? The agent needs to think (reason) and do (act).
The iteration part is fascinating. Remember having to manually rewrite the prompt in the workflow example? An AI agent could potentially handle that itself. It might critique its own output against certain criteria (like “LinkedIn best practices” or “brand voice guidelines”) and refine the draft multiple times until it meets the standard – all without direct human intervention in that loop.
Andrew Ng’s demo mentioned in the video (finding a skier in footage) is a great practical example. Instead of a human manually tagging “skier,” the agent reasons what a skier looks like, acts by scanning the video, and delivers the result. It handles the task autonomously based on the goal.
Why Does This Matter for Us?
Understanding this progression from basic LLMs to workflows to agents is really helpful, especially for professionals and creatives.
- LLMs are powerful assistants for generating text, brainstorming, and basic tasks with a specific prompt I wrote. Having a go-to set of instructions can be a huge timesaver; I actually shared some of my practical prompts for daily writing and social media content before.
- Workflows allow us to automate multi-step processes using AI, connecting different tools and data sources based on rules we set. This is where I have experience with marketing automation and tools like Looker Studio for reporting comes into play – setting up systems to execute defined tasks efficiently, reviewing excel sheet, confirm the data are clean or not (small scale analysis) etc.
- AI Agents represent the next step: AI systems that can potentially take on complex goals, figure out the steps, use tools, and adapt their approach autonomously. Imagine an agent not just running the ad campaign you designed, but analyzing performance, reasoning about market shifts, and adjusting the strategy itself to meet a defined goal like maximizing ROI or reaching a specific audience segment. This evolution naturally raises questions about the future of various roles and disciplines, including SEO’s relevancy and how it’s adapting in the age of AI.
While truly sophisticated, autonomous agents doing complex professional tasks are still evolving, the underlying concepts are becoming more accessible. We’re seeing simpler agent-like features appear in various tools.
Wrapping Up
I hope walking through this video’s explanation from a reader’s perspective has helped clarify things. The journey from simple chatbots (LLMs) to rule-based automations (Workflows) to goal-oriented, decision-making systems (Agents) is a significant one.
The key takeaway, as the video highlights, is the shift in decision-making. Workflows follow human instructions; Agents figure out the instructions themselves to meet a goal.
This field is moving incredibly fast, and understanding these foundational concepts helps us cut through the hype and see the practical potential for our own workflows, whether in marketing, content creation, business operations, or any other field.
What are your thoughts on AI agents? Have you experimented with any tools that have agent-like capabilities? Let me know in the comments below or connect with me on LinkedIn! I’m always learning and curious to hear about others’ experiences. Have you started exploring AI agents or workflows? I’d love to hear about your experiences or any questions you have in the comments below!
