Artificial intelligence (AI) has long ceased to be a topic of science fiction. Our lives were divided into before and after when OpenAI introduced ChatGPT (Generative Pre-trained Transformer) at the end of November 2022. Generative AI, such as ChatGPT, is capable of creating any text, photo, video, audio, and other types of data based on the information it has been previously trained on
This breakthrough was only the beginning of a big wave of changes. At the end of last year, a new trend related to AI started to gain momentum: AI agents burst into our lives at an incredible speed. Everyone is talking about them everywhere, just as it was the case with generative artificial intelligence.
Content
Are chatbots and AI agents the same thing? No, they are not. Chatbots are simply «answering machines». You ask a question — they answer it. But their work ends as soon as they complete the task. Autonomous AI agents, on the other hand, work towards a more complex goal. Imagine that you can hire an AI agent to learn instead of you: learn a new language, analyze your finances, or even play video games for you. And spice it all up with a virtual copy of you with your voice.AI agents were the hottest topic at CES 2025.
For example, Nvidia introduces AI partner to help PUBG players.
The company also announced launch of a personal AI supercomputer called Project Digits in May. The heart of Digits will be a new GB10 Grace Blackwell chip which has enough computing power to run complex AI models, yet is compact enough to fit on a desk and run on a standard outlet. Digits can run AI models with 200 billion parameters, and its starting price is $3 thousand.
Manufacturer of gaming equipment and peripherals, Razer presents Project Ava. This AI agent will turn an average gamer into an imba: for example, it will help with complex puzzles, bosses, and quests.
Technology startup Natura Umana showed Humanpods headphones — a wireless headset that gives users access to artificial intelligence assistance. The company’s Nature OS operating system allows users to communicate with their own LLM-based agents called «AI People» by voice
Based on a user’s verbal request, Nature can assign different AI People to address a specific need: acting as a therapist, fitness trainer, or tourist guide. Users can also access other LLMs, such as ChatGPT, Gemini, and Claude, without having to open any apps or even touch their mobile phone.
AI agent Runner H promises automate complex, cumbersome, multistep tasks without repetitive and manual input.
The ideal AI agent is described as a system that can perform a wide range of tasks, just like a human assistant: plan a vacation based on budget and preferences, buy all the necessary tickets, make reservations, add reminders and make a list of necessary things, and record each step in your calendar. Or, you can analyze a list of work tasks and take on some tasks yourself, such as composing a text and sending out invitations, reminders, or emails.
Also, a true AI agent must be multimodal, i.e., it must process speech, audio, and video. For example, in Google’s Astra demo, users could point a smartphone camera at objects and ask the agent questions. Astra responded to incoming queries with both text and audio and video.
An advanced AI agent could, for example, serve as an advanced customer service bot. Today’s bots based on large language models (LLMs) can only generate the next likely word in a sentence. But a truly autonomous AI agent should be able to respond to natural language commands and perform customer service tasks without the need for constant supervision.
For example, an AI agent will be able to analyze customer complaint emails, check their ID number, access databases such as customer relationship management (CRM) and delivery systems to check whether the complaint is legitimate and process it in accordance with company policy.
Princeton University researchers believeThe company says that current AI agents must meet three criteria:
Broadly speaking, there are two categories of AI agents: software agents and embodied agents
Together, these components give the agent the ability to work independently, performing tasks through a «trial-and-error» method to achieve the goal.Imagine that you gave the agent the task «to find useful information about Facebook». In this case, here’s what it will do:
«We are still a long way from having an agent that can automate all the work for us. Current systems «hallucinate and also do not always follow instructions precisely», — said Fan
Another limitation is that after a while, AI agents «forget» what they are working on. AI systems are limited by their contextual windows, i.e. the amount of data they can take into account at any given timeFor embodied agents such as robots, there are even more limitations. There is a lack of training data to train them, and researchers are just beginning to utilize the capabilities of fundamental models in robotics. So, despite all the hype, it’s worth remembering that AI agent research is still in its early stages, and it will likely take years before its full potential is realized.
Oracle Company allocates the following types of AI agents that can be used in various fields:
Creating an AI agent can be a challenging but exciting task. You don’t need to have deep programming skills, but you still need to have at least a basic understanding of how artificial intelligence algorithms work and how to develop them for a specific use case. Creating an AI agent requires a combination of technical and strategic skills. In traditional conditions, skills are required to create AI solutions:
A minimum of skill is required only when using no-code platforms, such as Relay.app, Bubble.ai, Voiceflow or Tars.
Marharyta Lanhenbakh, PhD, Senior Data Scientist:
«Intelligent agents can be used to describe a wide variety of things, from bots in customer support to assistants like OpenAI’s Operator. In general, an agent is a system that can independently perform complex tasks and make decisions using available tools to achieve its goals.
Effective agent implementation requires several key components:
Model. The basis of the system is the chosen model. In modern practice, these are mostly generative language models (LLMs, such as GPT), but other approaches can be used. For example, highly specialized models for selecting the optimal solution among several options (for example, an autonomous patient management system in a hospital; such a system monitors information about the patient’s condition and, if necessary, makes a decision to remind him or her to take medication or call an ambulance). It is important to take into account the specifics of the task in order to choose the model that best meets the requirements. Sometimes systems without models, where behavioral options are prescribed by an algorithm, are also considered agents.
A set of tools. The tools that an agent can use provide its functionality. These can be:
out-of-the-box solutions: no-code or low-code platforms that offer basic functions and templates for quick customization. Nowadays, a certain limited set of functions is often already available directly through the APIs of language model providers (OpenAI has a section on how to query using tools).
Custom tools: specially developed scripts or programs for non-standard tasks. For example, if an agent is required to automatically search for jobs on LinkedIn and send out resumes, the developer creates a set of scripts to collect data and automate the process.
Also important are Correctly formulated promotional messages. If the agent is based on a language model, it is important to create accurate and clear instructions (promt) that define what the system should do in each specific situation. High-quality promts allow the agent to perform tasks the first time without unnecessary clarifications or errors.
Building an effective agent is a balance between technical infrastructure, tools, and precise instructions. This allows not only to achieve the desired result, but also to ensure the flexibility and adaptability of the system to new tasks».
Let’s take one of the simplest examples: creating a personal AI assistant for communication on Telegram using the GPT3/4 model N8N — a good option for beginners, as you don’t need to have a programmer’s background. To do this, use Telegram trigger and Telegram node.The process consists of three stages:
After the bot is registered, work on its functionality begins. If you create an AI assistant on your own, the financial investment will be insignificant. The main thing is to make the bot available 24/7. This means that it needs to be hosted somewhere. A simple version of an AI agent will cost about $5 per month for a low-level VPS. If the bot becomes extremely popular, you will have to consider the need for additional computing power or pay to a commercial bot creation platform.
To do this, we take the following steps:
Next, we move on to creating an AI bot that can naturally respond with emojis, handle unsupported commands, and provide error messages. Our bot will also be able to create and send images generated by DALL-E 2.You can choose the cloud version of n8n (the price starts at $20 per month). To do this, register on the OpenAI platform and get a new API key. If you decide to host n8n yourself, don’t forget to set the `EXECUTIONS_PROCESS` environment variable to main.
Let’s divide the bot’s workflow into three main parts. The first part receives incoming messages from Telegram and performs preparatory actions:
This information allows the bot to greet users by name and take into account their language of communication.
Variable bot_typing is used to show a message that the «bot is printing a» response while it is working on the request. The message depends on whether the user started their message with the /image command (to request an image) or not. This adds interactivity and gives the impression that the bot is actively working.
Model_temperature stores a value «temperature». The higher the value, the more «creative«answers the model generates. For example, a low temperature provides more accurate and predictable answers, while a high temperature adds variability.Variable token_length limits the length of the response generated by the GPT model. This helps to control the amount of text sent to the user and avoid overly long responses.
Action Send typing uses the value of the variable bot_typing with JSON to show the user an animation that the bot «prints» response. You can also see the Merge node. It’s configured in ChoseBranch mode to pass the original JSON from the settings to the next steps. This is a simple technique that ensures that the Send typing action completes before the next steps start.
At this stage, the bot processes the user’s data and sends it to one of three generative artificial intelligence models, depending on the type of request. That is, GPT (if it is text) or DALL-E 2 (if it is an image)The bot analyzes the user’s request to determine which model should be used. For this purpose, it uses Switch node (CheckCommand) which redirects requests to the appropriate model. Setting up the Switch: the first three routing rules are related to correct requests and are transmitted to either the GPT or DALL-E 2 model. The last rule — is a backup option for unsupported requests.
Examples of routing:
.
After the OpenAI models return a specific response (text message or image), we need to pass this information back to the user. In addition, we need to prepare a generic response for unsupported commands. As you can see, Text Reply is connected to both OpenAI nodes. This is a small trick in n8n to reduce workflow redundancy, Send Image is transmitted via the URL returned from the Create Image.
Finally, Send error message returns a generalized response. Thus, we do not call any OpenAI model and provide an immediate response, such as an error message.
You can also use the n8n to create a personal assistant for managing your calendar and sending emails. Details video instruction here.
AI agents are actively changing the world by integrating into workflows, transforming business and entertainment. About OpenAI predicts the emergence of the first AGI agents that can perform complex tasks autonomously. Satya Nadella, CEO of Microsoft, said.
The company believes that traditional business applications will give way to integrated platforms powered by AI.
AI agents are already transforming gaming by acting as learning partners, allies, or NPCs with procedural content generation. For example, in open-world games such as Red Dead Redemption 2, NPCs remember previous encounters with the player and react accordingly, creating an exciting dynamic. In addition, AI agents are capable of high-quality procedural content generation (terrain and landscapes, quests and missions, items and loot, character design, etc.).
No Man’s Sky’s artificial intelligence creates entire universes with unique planets, creatures, and ecosystems, offering almost limitless possibilities for exploration AI agents can analyze player behavior in real time to dynamically change the game’s difficulty. Resident Evil 4 uses adaptive difficulty systems to fine-tune enemy behavior and item availability depending on the player’s level.
AI agents in the crypto market are now so popularthat in a few months the market capitalization of their related tokens has grown to $15.7 billion and, in forecasts CoinMarketCap, could reach $250 billion by the end of 2025.In Web3, AI agents enable decentralized asset management, trading, and the creation of «smart economies«.
They analyze trends, offer transparency, and are financed through tokenized models.For example, Degen Spartan AI combines social media data with market trends to open up new business models.
No-code platform AlchemistAIapp already allows you to create AI agents on your own. And the multi-agent coordination protocol Questflow increases productivity by integrating the capabilities of multiple agents.
Virtual influencers, such as Miquela Sousa engage millions of social media followers by generating content and interacting with the audience. Such agents can automate marketing campaigns, analyze sentiment, and improve their performance.
AI agents personalize learning processes, develop individualized plans, and provide access to real-world simulations. In healthcare, they support diagnostics and patient monitoring via IoT and help doctors make decisions. In marketing, they automate targeted campaigns based on user behavioral data.
AI agents provide autonomous transport management. For example, the new Mercedes CLA with the next generation MB.OS operating system has received an updated MBUX virtual assistant from the Automotive AI Agent platform from Google Cloud. Virtual assistants such as OpenAI Operator, perform tasks such as making reservations, filling out forms, and ordering products by integrating into web environments.
In business, AI agents analyze data, optimize processes, and offer multi-agent solutions that work on common goals such as logistics or production plans. In HR, agents automate recruitment, training, and onboarding, and help job seekers find the perfect position, as does Robin by Amply.
The future of AI agents.
By 2025, multimodal agents are expected to develop that will work with text, images, and audio. Their integration with IoT will ensure the control of smart homes and industrial equipment, and the emergence of AGI will allow the creation of systems that understand the context and learn on their own.
AI agents are changing the world around us at an unprecedented rate. They make life easier and open up new opportunities, but they also raise questions about ethics, security, and the impact on human society. Virtual friends, financial advisors, lovers, and even psychologists… But along with the prospects come serious ethical, security, and social challenges.
An AI agent or assistant is a complex program. And like any program, it can be hacked. In 2024, there have already been cases when virtual assistants have become loopholes for hackers. In addition, companies can use your data for purposes you don’t even know about (e.g., targeted advertising or selling to third parties). For example, Amazon admitted that Alexa saves user conversations. Therefore, it is important to use services that offer data transparency and the ability to delete it (e.g, GDPR-regulated). Regularly check the privacy settings in your apps.
AI agents are capable of performing many tasks, but still have technological limitations that are difficult to overcome. However, it is expected that by 2030, about 30% of routine professions (cashier, administrator, graphic designer, call center operator) will be automated by AI. 41% of global companies plan to reduce staff, writes CNN. On the other hand, it will lead to the emergence of new professions such as «AI trainer» or «data architect». Therefore, it is important to keep abreast of technology, as there is a risk of being unemployed.
Ready to pay for an AI friend, psychologist, or lover? Recently, more and more services have appeared where people pay for emotional support, friendship, or even virtual «romantic«relationships with AI. This trend is a matter of controversy. On the one hand, AI psychologists such as Woebot or Replika, can help for people with social anxiety, depression, or stress, providing 24/7 support. Such a service will cost less than a real psychologist. And it is available at any time. However, it can cause emotional dependence in the user. Besides, AI can not fully understand human emotions and may offer standard or even harmful advice in a particular case.
Humans can «train«AI in toxic or even dangerous behavioral scenarios, which raises issues of control. There is a risk that people will start avoiding difficult conversations or conflicts, communicating only with «ideal» AI friends and lovers. And the mean ones will start losing their socialization skills altogether.
Another important aspect is the psychological impact on users. Communication with AI can create the illusion of intimacy or deep connection, when in fact it is just an algorithm programmed to mimic human reactions. People may begin to perceive virtual relationships as equivalent to real ones, leading to emotional isolation.
