AI Agents: The Next BIG Revolution in Artificial Intelligence!

Vídeo em Inglês, publicado pelo canal Nishant Chahar.

About this content: The summary, key points and topics below are generated automatically from the video transcript. They reflect the content, opinions, claims and predictions presented in the video and do not represent a position, validation or fact-check by IA Dots.

Video summary

The video explains the concepts behind large language models, workflows, and AI agents, detailing how to build a basic AI agent and highlighting the differences between traditional chatbots, workflows, and autonomous agents.

Key points

• Large Language Models (LLMs) function primarily by predicting the next word based on training data and are inherently passive.
• Workflows connect LLMs to external sources or documents using control logic and Retrieval Augmented Generation (RAG) or tool calls.
• AI agents differ from workflows because the AI decides its own steps and path based on the output it receives to achieve a specific goal.
• The core framework of an AI agent involves reason, act, and iterations (loop engineering) to refine outputs autonomously.
• An AI agent consists of four main parts: the LLM (brain), tools, memory, and a loop.
• Users can build simple agents using platforms like 10810, Zapier, or Claude by providing a clear goal, instructions, and tools.

Topics

AI AgentsLarge Language ModelsWorkflowsRAGReason and ActLoop EngineeringEvaluations

Reading time for this section: about 4 min · Video length: 11 min 8 s

Jump to

• Why this explainer is worth your time
• Our take (interpretation, not fact)
• Facts you can check
• Tools and ideas mentioned
• More on AI agents at IA Dots
• Your turn

In 30 seconds

• An LLM predicts the next word, only knows what it was trained on or given access to, and does not act on its own.
• A workflow connects the model to calendars, documents (RAG) or tools along a path you define; an agent picks its own path toward a goal.
• The video’s four parts of an agent are the model, tools, memory and a loop, and it says evaluations matter more as the agent’s tasks get riskier.

Why this explainer is worth your time

“AI agent” gets attached to almost every product launch now, which makes it hard to tell a real agent from a chatbot with a plug-in. This video gives a simple test you can reuse: if you decide every next step, it is a workflow; if the model decides the next step based on what it just got back, it is an agent.

Our take (interpretation, not fact)

This is IA Dots’ interpretation, not a fact. The most useful part of the video is not the definition but the warning at the end: start with low-risk tasks, add guardrails, and check outputs by hand before trusting an agent with API keys or real actions. That advice travels well anywhere, because the cost of a wrong action grows faster than the time an agent saves. The workflow-versus-agent line the video draws is close to the one Anthropic published in its own engineering guide, which suggests it is becoming a common vocabulary rather than one creator’s framing. On cost, you can try the ideas without paying: Claude has a free tier, n8n offers a free self-hosted Community Edition, and Zapier lets you get started free, while heavier use moves to paid plans. Prices below were checked on Oct 31, 2026 and may have changed. Note that the auto-captions write n8n as “10 8 10”.

Facts you can check

• Anthropic, “Building effective agents,” Dec 19, 2024 (checked Oct 31, 2026): Anthropic defines workflows as systems where LLMs and tools follow predefined code paths, and agents as systems where LLMs dynamically direct their own processes and tool use.
• arXiv paper 2210.03629, submitted Oct 6, 2022 (checked Oct 31, 2026): the “reason plus act” idea the video calls the ReAct framework comes from the paper “ReAct: Synergizing Reasoning and Acting in Language Models” by Shunyu Yao and six co-authors.
• n8n pricing page (checked Oct 31, 2026): the Starter cloud plan is listed at €20 a month billed annually for 2,500 workflow executions, and a standard self-hosted Community Edition is available on GitHub. Claude’s pricing page lists a $0 Free plan and a Pro plan at $17 a month billed annually ($20 monthly). Prices were checked at that moment and may have changed.

Tools and ideas mentioned

• n8n (written “10 8 10” in the auto-captions)
• Zapier Agents
• Claude
• ChatGPT and Gemini
• RAG (retrieval-augmented generation), tool calls and MCP
• The ReAct pattern, memory files and evaluations (evals)

More on AI agents at IA Dots

• Meta Muse AI Email and Calendar Automation Explained | Beginner's Guide
• Here Is My ENTIRE AI Setup (Nothing Held Back)
• The Personal AI Shift: OpenClaw, Muse, Instinct, and What Comes Next with Brad Groux
• What Happens When AI Runs the Robotics Experiments? - Jia Qi Yip

Your turn

The video’s starter project is a daily news agent that searches the web and sends you the five biggest AI stories at 8 a.m. If you built your first agent this week, what single low-risk task would you hand it, and what would you still check by hand? Tell us in the comments and share this post with someone who keeps asking what an agent actually is.

This summary was generated with AI assistance and reviewed by humans.

Video transcript

Transcript from available captions; it may contain errors.

Right now the people who are getting most out of AI, I don't even know how to actually do it. But they were able to figure out one thing, how to make AI work for you. And that is exactly what AI agents do. In this video, in the next 10 minutes, I'll teach you how you can build your own first AI agent, and also tell you what is the difference between a normal chatbot that you use everyday, which is ChatGPT or Claude, and an AI agent. So instead of just asking questions to ChatGPT, you will be able to give an agent a task. It will go through multiple steps and give you the final output. So stay with me till the end of this video. Hi guys, I'm Nishant Sahai. I'm an ex-Microsoft software engineer. I've built and sold my first startup average around 250k US dollars for my second startup, and I am building multiple products right now. So without any further ado, let's start the video. Let's first understand what an AI agent actually is. [music] So AI agent is actually a system. You give it a goal and it figures out all the steps in between to finally give you the final output for that particular goal. It will figure out all the tools, it will figure out what needs to find, it will figure out all the steps on its own. Now to understand this properly, we'll go through three steps to first understand what an LLM is, then what a workflow is, and finally what an AI agent is. So whenever someone asks you what an AI agent is, you'll be able to explain that person like very simple terms. Let's first understand the foundation. Now if you want to understand what an AI agent is, you need to first understand what LLM is. LLM means large language models. ChatGPT, Gemini, Claude, all of them are tools built on top of large language models. Now the question that comes in your mind would be what is the work of an LLM? So what an LLM does is it mostly predicts the next word. It is trained on a large set of data and it learns how to predict the next possible word, and it will generate something for you. That is why it's called generative pre-trained transformer. For example, I asked ChatGPT to write an email for going on a coffee chat with someone. So my prompt is going to be the input, and the output that I'll get is the email. But if I asked ChatGPT when is my next coffee chat, if it doesn't have access to my calendar, if it doesn't have access to other tools, it will not be able to answer this question. The reason is ChatGPT doesn't have the calendar access, doesn't have the email access, and it can only give you output on the things that it has learned previously. And from here we are able to learn two main traits of LLMs. LLMs only know the information that they are trained on or they have access to. And second, LLMs are passive. They won't be able to take actions on their own. You'll give them some prompt and they'll give you some answer. Just remember this because we'll be solving these limitations in the next step. And the next layer is called workflows. So, if you remember in the last layer we asked ChatGPT, "Can you tell me when my next coffee date is?" And it was not able to answer us because it did not have access to Google Calendar. Now, the question is can we connect [music] these LLMs to external sources? Now, it is even more easy because they have added plugins and other things. And now in general ChatGPT itself acts as an agent, but if you go back a few months, this was not possible. And if you think about this question, if you had to give access to external sources to your LLM, how would you do that? And from here the AI workflow concept actually starts. For example, I just tell LLM, "Whenever I ask you about a personal event, you have to go to Google Calendar and find if there is something relevant for that." So, the next time when I ask the LLM that when is my next coffee chat, so the LLM is going to first check the calendar and then answer me with the exact date and time. Now, think if there was a follow-up question in this chat [music] itself where I'm asking, "What would the weather look like on that day?" It will not be able to answer because it's just going to the calendar and checking [music] for personal information. And this is the core trait of an AI workflow. It will just follow the process that we have defined for it. And technical language we call this control logic. And the thing that you just did, giving external sources to LLM so that it can get information from there, this is a very common idea. When we get this information from documents or knowledge bases, this exact thing is called RAG, Retrieval Augmented Generation. Just think of it. There is a 200-pages document related to your company's HR policies. And if you're asking some question related to that, "That how many sick leaves do I have?" The normal LLM won't have an answer to this, but if you use RAG and give the context from these 200 pages, it will actually find the exact page, add that to the context, and then give you the answer. Our calendar example is a little different because we are not giving them any document. We have added the access to Google Calendar here and that is called a tool call or an MCP call. You just need to remember that rag is not a different system. It is just a way to provide information to an LLM so that it can take decisions and give you answers. Now you need to notice one thing. Whether your workflow has five steps or 500 steps, until we have already decided what the next step is going to be, it is going to be a workflow. Because I'm doing the actual trial and error, I'm able to figure out what should be the next possible step. What if AI is able to decide what the workflow should be, what should be the next possible step depending upon the output that it is getting, that is going to be an AI agent. Whenever AI is deciding what should be the next possible step, that is what an AI agent is. And here comes the actual difference between an AI agent and an AI workflow. When you're building a workflow, you're actually deciding the next steps. But when you're building an AI agent, you just give it a goal and AI agent or the LLM behind it will decide the next feasible steps. For example, you want to create a LinkedIn post based on some articles. Currently, I will have to decide which articles do I want to read or which articles I need to add so that I can get the LinkedIn post. I'll have to figure out what are the relevant tools, what is the relevant information, and what should be the next possible steps. If the agent is taking these decisions using LLM, that will change the whole picture. And we have to understand these three important things. The first one is reason. Reasons mean that you have to figure out from the next possible options what you have to pick. For example, agent needs to think if I have to collect information, what are the best possible resources for this. And for that, what are the tools that I will need to call? Like whether I have to find a tool that can help me search or do I use the data which I already have? Second is act, which means if the agent has already decided, then using that tool or using that API, doing the actual work is called act. And for this particular reason plus act is actually called the react framework. And the third trait is iterations. For example, if you don't get the right output in the first attempt, it will have to figure out and reiterate and find the best output. If you remember, if the output is not correct in the workflow, I'll have to go and change the output and rerun the whole workflow. Now here I don't have to do that. The agent can actually critique the output itself and then figure out whether it is up to the mark or they have to do run the whole thing again. What the AI agent will do is it will regenerate the improved output, then reassess whether it is good or not, and if it is good, it will give you the final output. Otherwise, it will go through the whole iteration again. This is done until the goal is completed. The same thing is also known as loop engineering. If you want to know more about that, do let me know in the comment section. And this is the major difference between a chatbot and an agent. Let's understand this with a simpler analogy. Either you want to work in a big tech company or you want to work in a startup. What happens is there is a senior person, a senior engineer, who will decide what task you want to do. They will just give you the task. You will have to figure out what the steps need to be. You will go through the steps, you will do everything. If you have the perfect output, you will report back to the manager or the senior engineer, whoever has given you the task. Now, there are four core parts of an AI agent. Let's understand those. Now, you define a goal for an agent, it will reiterate until the goal is completed. Now, let's understand what exactly is inside an AI agent. Basically, an AI agent has four main parts. The first one is LLM. This is the brain of the agent. This will help it decide what are the actions it needs to take, what are the steps it needs to do, and also help the agent evaluate whether the task that has been given and is being performed by the agent is correct or not. The second is tools. These are the actual tools, like if you are a mechanical engineer or if you're an engineer and your car breaks down, you will have some tools to fix that car, right? Similarly, the AI agent has some tools, like it may be an MCP, it may be a search API, it may be another LLM's API key, things like that. [music] Third is going to be the memory. This means agents will remember what they have done. The memory is mostly saved as a markdown file or a dot md file. Whenever they make a mistake, they will add that to the memory itself so that they don't repeat the same mistake again. And the last one is the loop. This process will keep on going until the final output is there. The same thing I use for my video editing agent that I have created, which edits a lot of my videos. So, LLM, tools, memory, and loop. These are the four things that make an AI agent what it is. Now, the question is, can you also make these AI agents? The answer is yes, and you don't even need to know how to like write code. There are multiple platforms like 10 8 10, there's Zapier, even Claude. You can use any of these platforms to create your first AI agent. Now, let's understand this in simple terms. The agent needs four things: a clear goal, clear set of instructions, that what exactly a good result looks like, tools, and the freedom to actually go through this again and again. For example, you can create a daily news AI agent which will go scrape information from Twitter, from Reddit, and find you the daily news that is happening in this world. You can just give the goal like, "Find the five most important AI news stories from the past 24 hours. Give a short summary of each and explain why each one matters." Now, give it a tool like a web search, and just schedule this for 8:00 a.m. every single day. And that is it. The agent will go on, do the research, and give you the final output. And as I told you, you can use 10 8 10, Zapier, or Claude to build the same agent. And one thing that I wanted to tell you is, keep your first agent as simple as you can. It's just to understand what an agent actually looks like. The more freedom that you give, the more tools that you give, it will be more difficult for you to actually use that agent and and manage that agent. And one important thing is, whenever you give task to an AI agent, giving 100% full freedom is not the right thing always. If the agent is just doing browser surfing, researching, creating a document for you, the risk profile is very less. But if you're creating a product, you're giving API keys, you're doing XYZ things, it might be very harmful for you. That's why start small, give low-risk tasks, understand how everything works, and then you can improve the risk tolerance for yourself, and also add some guardrails so that your agent doesn't go haywire. And whenever the AI agent is actually doing a high-risk task, or where it is taking a real action, a review is needed. This is a very important concept that you should know about. It's called evals or evaluations. The name sounds very technical, but the concept is very simple. It's what we have been doing throughout our life. We learn something, and then we get evaluated on that. Because AI agents sometimes will give you wrong answer confidently. To stop that, we have evaluations. And there is also a possibility that it is just regenerating multiple outputs for the same task, which is just wasting time as well as tokens for you. So, without testing, how will you be able to figure out whether your AI agent is right or not? And is it genuinely reliable for you or not? And that is why you have to regularly check whether the output is correct or not, and whether there is an issue or not, and if the agent is working reliably, plus it is performing as per the expectations that you have. As the agent does more important tasks, the importance of evaluations also increases. Because you need to also know whether the agent is working properly or not, and when it is actually failing, and why is it failing? That's why you should never trust agents blindly. You should manually check the output so that you actually know how reliable your agent is. This is all about AI agents that you need to know. I hope you liked this video, you were able to understand what an AI agent is in very simple terms, and you can automate a lot of your daily tasks that you do manually by AI agents. Do let me know in the comment section which AI agent you are going to build now, and don't forget to subscribe. See you in the next

Assistir ao vídeo original no YouTube

Assuntos: Agentes de IA, Assistente pessoal. Publicado no YouTube em 2026-10-30.

Seleção automática por assunto e idioma. O vídeo pertence ao canal de origem.