Что Такое ИИ Агенты — Полный Гайд Для Новичка

0:00Friends, hello everyone. So, everyone's talking about agents these days. But let's be honest,

0:05few people even understand what they are or how to actually assemble them. You watch videos, and they

0:11talk about agents as if they have their own little brain inside, thinking and reasoning on their own.

0:16In short, it's pure magic and hype. So let me give you the honest version in one sentence.

0:21An agent is a language model that can call external and internal tools

0:28and work in a loop to perform tasks until they are considered successfully

0:34completed. And all this is what a modern IAgent is. And in this video, we'll

0:40figure out what IAgents are without all the fluff. We'll see what's really going on

0:46under the hood. We'll assemble the same agent in completely different ways. From completely no-

0:51CD solutions to writing an agent in code. By the end of the video, you'll understand this whole topic better

0:56than 99% of people. Subscribe below. Leave a thoughtful comment, give it a like, and off we go.

1:04First, let's clear up the most basic confusion. What's the difference between chatbots and agents? A regular chatbot,

1:11or simply a chatbot, can only do one thing: chat. You message it, it responds. That's where the interaction

1:18ends. It won't check your email, Google anything, or book anything;

1:24it simply suggests text. An agent, however, is essentially what you get when you empower your

1:31chatbot with additional capabilities. The first is the ability to pull levers in the real world. We call

1:36this lol calling, or calling up tools. The second is the ability to move forward independently,

1:41step by step, rather than simply stopping after one answer. In other words, you ask the bot a question,

1:47and you set an agent a goal, and it actually goes about achieving it. Once we understand this,

1:53all the magic essentially disappears. Let's break down the basics. The first is LLM. The Large Language

2:00Model is our brain. At the center of any agent is a language model. But let's be clear:

2:06all it does is predict the next tokens based on the previous ones. It has no arms

2:12or memory. It's basically sitting in a box, generating text. So how

2:18can it do anything useful? That's where to lol collin comes in. It's about its

2:24arms. You tell the model in advance. Look, I have these tools, like web search.

2:30Nero never magically goes online. It generates structured text

2:36that tells your shell: "Hey, I want to pull this tool, this instrument," and your code

2:43that runs this model actually goes online, gets the results, and feeds them

2:49back to the model. Passing this data back and forth is how agents work at scale.

2:55Next. The context window or Window context. This is the agent's short-term working memory. The model needs to

3:02somehow track what's going on. Everything gets into the context: your instructions,

3:07chat histories, the results of tool calls. But there's one caveat here. It's not elastic. You can't just

3:13dump thousands of page documents there. That's why context engineering is super important here,

3:18the ability to skillfully throw into the window only what is needed. And all this lives in the message structure.

3:25The system prompt is where we set the rules: "You are such-and-such a researcher, always specify

3:31sources." User prompt is our request or a classic one. prompt. And the assistant is what

3:38the model responds to. And remember a critical thing: the model itself has no memory. If you don't feed

3:44the entire conversation history back into the context each time, it will forget everything. And what

3:49if the history has grown and the context no longer fits it? This is where a vector database most often comes into play.

3:55We take our documents, turn them into embeddings, vectors, and give agents tools to search

4:01this database. This is our beloved RAG Retrieval Augmented Generation. This is how agents remember long-term

4:08facts and work with our knowledge base. And the best part is The Loop. This is where it all comes

4:15together. This is what many people miss. You give a system prompt and a goal. The model thinks, "Aha,

4:21we need to call the tool." Your code calls the tool, returns a result. The model

4:27looks at the result and thinks, "So, what do we do next? Should I call another tool or am I already finished

4:33and it spins in this cycle, that is, thinking, acting and observing, until it achieves the goal.

4:40So, that's the entire agent cycle. So, theory is cool, but how do you actually assemble agents?

4:45Let's divide them into four basic types. The first is no-code. No code, just forms or

4:51questions. The second is low-CD, visual editors, but the third is shells or agent harnesses.

4:59Powerful shells are ready, where we embed skills and tools. And the fourth is food, meaning we

5:04write everything from scratch. Totally hardcore. Friends, if you've watched this far, it means you're interested.

5:09Please like, subscribe, and comment on this video. Some kind of informed commentary. If you're

5:14just starting out in artificial intelligence, be sure to drop by our pro group. We have

5:19a bottomless amount of pro materials that cover a huge number of aspects of working

5:24with artificial intelligence, as well as ProChat, of which I'm a member. And master classes on Cloud

5:29CodeD, Envoy Mmen, and Google, so come visit us, the link is in the description. Let's get building. Let's start

5:36from the top. No code. There are actually a huge number of such tools. Today we'll

5:41look at Jнes Park. There are also Teeki, Gamloop, and a huge number of similar ones. Their goal is simple:

5:49they preserve our familiar experience. Essentially, they're a chat interface, but they're enriched with a huge

5:54number of connectors so you can, for example, integrate various tools, your

5:59Google Drive data storage, email, add some advanced capabilities, and skills so

6:05that a quick agent can be generated for your question, which would then execute it and return

6:10the result to you, for example, and then save it to your Google Drive. I say: "Let's create an agent

6:15that will first do research and tell you about the top IT companies in the world in 2026

6:20, then save it to my Google Drive. First, it does in-depth research,

6:25then selects the right agent and continues to do all these things with literally zero setup.

6:30And now we see how it integrated with different tools, did research, said that

6:35Nvidia is now the most expensive company, and then below it offered me to save it to my Google Drive.

6:40It's easy to start, simple to use, but there is minimal control, minimal flexibility. Level two-code. We have

6:47looked at this type of solution many times. That is, it's just fire for those who already want

6:52to work more seriously, but don't want to delve into development yet. Examples: Flowwise, Lengflow and our

6:59beloved NVC, of ​​course. So, an example. I have automation here on NVC Mman, which does roughly

7:06the same thing as the previous agent. Such automations most often consist of ND, that

7:11is, of composite pieces that all work in A pair and, let's say, controlled by some agent.

7:17Here we have Neronka, which receives our requests and understands what needs to be called. There's

7:23the Perplexity tool, a resarch search for researching the issue, accessing the internet, and

7:29gathering context. By the way, in our pro group, we have a masterclass on Nman, which has already become a cult classic,

7:35and it generally takes people from superzero to a ready-made agent with Rago and connecting their own

7:41neural networks. So, be sure to drop by and take this masterclass too. Here we continue

7:46and say the same thing. Do some research, tell us about the top IT companies in the world in

7:512026, and then save it to my Google Drive. We send this question. We see how the agent

7:55works, we see how it pulls Perplexity to do the research.

8:00We can always look at taxes, what exactly our Perplexity returns, everyone, an answer is being prepared, and now

8:05it will be displayed back to us in the chat. World IT leaders of 2026: Microsoft and Amazon, Apple,

8:12by the way, Nvidia For some reason, it's only in fourth place, which is strange. Pereplexity, generally speaking,

8:17should do a pretty good resarch. Well, it's what it's made. Perhaps we're missing something.

8:21Next, if we connect our Google Drive here, we'll set up saving this

8:27research to Google Drive and further work with this data. Next. Level three. Agent harnesses or agent

8:33shells. This is when we don't build anything from scratch, but install a powerful, ready-made agent.

8:40Several leaders right now are Code, Hermes, Codex, and I can't help but mention it. ENT, of course. Let's

8:46look at NT. The operating principle is the same, but our goal here is not to write any tools.

8:53Loops. It already does everything for us. Our task is to add skills, prompts, and MCP servers.

8:59So, I open the Agent interface and, without any preliminary configuration, say the same thing.

9:05Let's create an agent that will first conduct research and report on the top IT companies

9:10in the world in 2026 and then save it to my disk as a text file in the

9:15research folder. A new project is created, many tools are called up, and the agent automatically determines

9:20which tools need to be used based on the context. By the way, friends,

9:24I usually put all the resources we visit in YouTube videos in my free Telegram group. So

9:29go there, browse the sources, and you can follow me through the video itself. In the meantime,

9:35we've created a simple agent for researching the IT market. The first research has been completed and even

9:40the first research has already been saved to a text file. We see this directory, open it, and see

9:45the file is here. We open the file and see that, among the top IT companies,

9:51Nvidia is again in first place, followed by Microsoft. Apparently, the agent was more accurate the first time than the second.

9:57Shells and agents are good because they require minimal configuration, and if they don't know something, they

10:03can configure it themselves, make a decision, and continue working until the task

10:08is completed. And finally, level four. Developing agents from scratch. This is already a story for

10:14developers. When visual designers and ready-made agent shells aren't enough,

10:20we start writing agents in code. The interesting thing is that almost no one writes

10:25agents from scratch these days. They usually use specialized frameworks. Probably the most famous is Lgraph

10:30from Longchain. It allows you to assemble agents as a state graph. That is, it thinks here,

10:36calls tools here, and decides whether to continue the cycle here. This is very convenient

10:41when the logic becomes complex and different execution scenarios arise. The second popular

10:46option is the Open AI Agents SDK. It's much simpler. You describe the agent, connect tools,

10:54define instructions, and everything else is just cycles, tool calls, and context passing. The library handles

11:00everything for you. And it's a great option for most projects. There are other

11:04solutions like Pidentic AI, Crew AI, Autogen, and Semantic Knal. Well, the idea is pretty much the same. They

11:12free us from having to write the entire agent infrastructure ourselves. So, full code

11:18today doesn't necessarily mean writing thousands and thousands of lines of code. It means

11:24we usually choose a framework, leverage its capabilities, and embed our applications to

11:30leverage its power within our own developments. This is where developers gain

11:35maximum flexibility. You can connect any databases, any API, your own tools, your own memory

11:41, and completely control how the agent operates. So, friends, let's sum it up. What should you choose?

11:48The approach is this: if you're not a techie and want results in 30 seconds, use code like the ones

11:54we've reviewed. If you need complex logic, branching, and connectivity to a bunch of services, then it's clear

12:01there are probably no alternatives. We'll use N8N—it's fantastic. If we need a powerful production agent

12:07that can be customized to the fullest, we'll use NT or any other similar framework.

12:13But if you're developing your own product and want to embed an agent directly into your app,

12:18then we'll look at Landraph, Open AI, Agents SDK, and other software frameworks. The point is,

12:24they're all the same under the hood—a LMCAK that can be enriched with external and

12:31internal tools and answer our questions. The only difference is how much control we want

12:37over it. Definitely try building your own agents or use ready-made ones. Links

12:43to the resources we discussed in the free Telegram group, a pro-community with master classes,

12:49skills, and other pro-materials, are also in the description. I bid you farewell, and until next time.