April 2026 | issue 4
Most people have already encountered ChatGPT, Gemini or Claude. We enter a question, receive an answer, and that is the end of the matter. Such a system can write an email, summarize a document, explain a difficult subject or prepare a travel plan. It still waits for a human, however. By itself, it will not open a mailbox at eight in the morning, check new invoices, save attachments to a disk or send information that it has detected a problem.
W-Moszczynski-Agent-AI-PZM-4-2026A chatbot answers. An agent acts
An AI agent works differently. It is not merely an interlocutor. It is a program that can receive a goal, plan successive steps, use tools and carry out a task. It can operate on a server around the clock. It can use email, a calendar, a browser, files, a messenger, a database or a company system. It can also return to the user with a result without waiting for another question.
The simplest example looks like this. An ordinary chatbot must be told: “Review these three messages and prepare a reply”. An agent can be given a standing instruction: “Every day at 8:00, check new messages from customers, mark urgent ones, prepare draft replies and send me a summary on Telegram”. A chatbot carries out a single conversation. An agent carries out a process.
The difference can also be seen clearly in the example of a calendar. A chatbot, such as ChatGPT, can suggest how to write an invitation to a meeting. An agent can open the calendar, check available dates, find a time that suits the participants, prepare the event and ask a human only for approval. This does not mean that an agent is an autonomous artificial human enclosed inside a computer. It is still software. Its ability to analyse instructions comes from a language model, and its ability to act comes from connected tools. When an agent has access to Gmail, it can read and send messages. When it has access to Google Calendar, it can check dates. When it has access to a server, it can run scripts. Without such connections, it remains mainly a chatbot with better memory.
What an AI agent consists of
The first component of an agent is a language model—its decision-making engine. This can be a model from OpenAI, Anthropic, Google or DeepSeek, or an artificial-intelligence model run locally on one’s own computer. The agent sends it a description of the task, the current context and the results of previous actions. The language model assesses what should be done next.
Suppose that the agent receives the instruction: “Find the latest invoice from the transport company and save its data in the spreadsheet”. The model may determine that it must first open the mailbox, then search for messages from the supplier, check the dates, download the attachment, read the invoice number and finally enter the data in the spreadsheet. The model chooses the steps, but the appropriate tool carries out each step.
The second component is memory. An ordinary conversation with an ordinary chatbot—our proverbial ChatGPT on the phone—resembles a meeting with a consultant who forgets details after a while. An agent can store information in files or a database. It can remember something it heard two months ago: that the user does not want meetings before nine, that invoices should go to a specific folder, and that a sales report should contain three agreed indicators.
The third component is tools. A language model by itself—or, more simply, ChatGPT by itself—does not send an email or book a meeting. It can only generate text. Only an agent equipped with an appropriate connection to email or a calendar performs a real operation. This is a difference similar to that between an adviser and an employee. An adviser makes suggestions. An employee opens the system and does what needs to be done.
The fourth component is the task-triggering mechanism. An agent can operate on demand, according to a schedule or after detecting an event. A schedule can run a report every day at 7:30. An event can be the appearance of a new message, file or customer request. Thanks to this, an agent does not have to wait for a human to open the conversation window.
The fifth component is safety rules. An agent should know what it is not allowed to do. It may have permission to prepare a draft message but not to send it. It may move files but not delete them without confirmation. It may book meetings only on working days and only between nine and five. Without such boundaries, even a correctly functioning system can make a costly mistake.
A simple working day with an agent
Imagine the owner of a small company. In the morning there are thirty messages in their mailbox. Some are advertisements, some are questions from customers, and several concern overdue payments. An ordinary chatbot will help only when the user copies the content of a message into it. An agent can check the mailbox itself, recognize the subject of each message, move spam, prepare replies and mark matters requiring a human decision.
At ten o’clock, the agent can retrieve sales data from the accounting system and compare them with the previous week. When it notices a decline in the number of orders, it can prepare a short explanation: “Sales fell by 12%. The largest decline concerned garden products. The campaign that ran the week before was not launched this week”.
Interesting why companies suddenly stopped looking for data analysts en masse?
In the afternoon, the agent can check the calendar and find a free date for a meeting with a customer. It can prepare an invitation but send it only after approval. It can also send it without approval. In the evening it will record the actions performed in memory and create a short summary of the day. It can conduct a mass mailing without asking; it can give us its action plan for approval. Everything just like a real employee.
Another example concerns a person looking for a job. An agent can check new messages from recruiters, extract the job title, rate, working arrangement and requirements from them. It then saves the information in a spreadsheet and prepares a reply. The user does not have to copy data manually from a dozen or so messages. Nevertheless, they still decide whether the reply will be sent.
This is not magic. Every step requires configuration. Email, the calendar and data must be connected. Rules must be established. It is also necessary to decide which actions the agent may perform independently and which require human consent.
OpenClaw: an agent based on ready-made skills
OpenClaw is conceived as a personal AI assistant that can be expanded with ready-made skills. Its strength is the large number of free add-ons. According to the study analysed, the Claw Hub catalogue contained more than 50,000 skills. Instead of building every integration from the beginning, the user can install a ready-made connection to email, a calendar, a drive, an image generator, a CRM system or another service.
This is convenient for a beginner. The user does not immediately have to understand code, protocols and API interfaces. They can launch a web panel, choose a function and configure access. In this respect, OpenClaw resembles a smartphone. The telephone itself is a platform, but its real capabilities appear after applications are installed from the Play Store.
OpenClaw’s behaviour can be described in ordinary text files. The “agents.md” file contains operating rules. It can state that the agent does not send messages after six in the evening or delete documents without confirmation. The “soul.md” file defines its manner of communication. It can establish that replies should be short, factual and free from excessively polite formulas. The “user.md” file stores information about the user, their habits and work. The agent can remember that the company owner does not organize meetings on Monday mornings and wants to receive financial reports in PDF format. In turn, “memory.md” can contain information about earlier conversations and completed tasks.
Such a solution is simple and specific. When the agent behaves badly, the appropriate file can be opened and its wording corrected. If it begins preparing excessively long messages, the style file can say: “A reply must not exceed five sentences”. The entire system does not have to be rebuilt.
OpenClaw can operate actively. It does not have to wait for a question. It can check new information according to a schedule and be the first to write to the user. At 7:45 it can send a message on Telegram: “You have three meetings today. The first begins at 9:30. There are two urgent messages from customers in the mailbox. I have prepared replies for approval”. It can also check the prices of selected products, the status of a shipment, the weather forecast or the condition of the company website. If it detects a problem, it sends a warning. A human does not have to remember to run the check regularly.
Convenience has a price, however. A large catalogue of add-ons means a large number of components that must be trusted. Every plugin can contain errors. A malicious add-on may try to steal data or perform an unwanted operation. A beginner may see a popular name, click “install” and fail to check who created the module. In the case of an agent with access to email and files, this is a serious risk.
Hermes: an agent that remembers the way you work
Hermes represents a different philosophy. Instead of relying primarily on a large catalogue of ready-made add-ons, it focuses on learning repeatable ways of working. An important component is the Curator mechanism. When the agent solves a complex task, it can transform the completed process into a simpler and cheaper skill for reuse.
Suppose that every week the user downloads a CSV sales file, removes empty rows, groups results by region and prepares a PDF report. The first time, Hermes can carry out many steps with the help of a language model. The model analyses the instruction, checks the data and decides what to do. This uses tokens. When the process begins to repeat, Curator can save it as a ready-made script. During the next run, the agent does not have to analyse every step again. It runs a proven procedure. The language model is used only when an unusual situation arises—for example, the file format changes or one of the required columns is missing.
This is an important difference. The language model to which the agent connects is more expensive than an ordinary script from ChatGPT. The instruction “remove empty rows from the address column” does not require advanced reasoning every time. Once the procedure has been saved, the computer can perform the operation quickly and cheaply. Cheaply, because it will require fewer tokens.
Hermes can also learn the user’s habits. This does not mean merely remembering facts. The agent can notice that the user always wants the report in the same layout, records dates in day-month-year format, and archives files in folders named with the week number. Over time, fewer and fewer things have to be explained from the beginning.
An example may be the analysis of job offers. The first time, the user describes precisely that they are interested in AI Engineer positions, remote work and a rate above a certain level. Hermes observes how the offers are assessed. After a while, it can prepare a table by itself, reject unsuitable proposals and mark those that meet the criteria.
Hermes is more technical to operate. Its natural environment is the command line, called the CLI. For someone who has never used a terminal, a black window with text may be less friendly than OpenClaw’s web panel. This does not mean that Hermes is intended exclusively for programmers. A beginner may, however, need help with installation and initial configuration.
Two different philosophies
OpenClaw says: “Choose a ready-made skill from a large catalogue”. Hermes says: “Show me the process and I will learn to repeat it”. Both approaches make sense.
OpenClaw will be the natural choice for someone who wants to connect several popular services quickly. When a connection to Gmail, Google Calendar, Drive and a messenger is needed, ready-made modules shorten the route. The user does not have to write every script themselves. Hermes will be attractive where a company has its own, repeatable procedures. An example may be the daily inspection of production files, weekly data cleaning, generation of a report in an agreed format or checking specific websites. The agent learns the workflow and over time performs it increasingly cheaply.
The difference resembles a choice between a large DIY store and a workshop. In the store, a ready-made component can be bought quickly. In the workshop, a tool can be created to fit one’s own process. The store provides speed and choice. The workshop provides fit. OpenClaw offers a greater choice of AI-model providers and ready-made integrations. Hermes concentrates on a smaller number of connections but develops the mechanism for independently preserving procedures more strongly. OpenClaw more often requires the user to approve a new skill. Hermes can optimize it automatically. There is no single winner here. A user who needs a graphical panel and quick integrations may feel more comfortable with OpenClaw. A user who accepts the terminal and wants to preserve their own procedures automatically may choose Hermes.
Memory, or why an agent does not constantly start from zero
Memory is one of an agent’s most important components. Without it, every conversation would begin with the rules being explained again. The agent would not know the names of customers, where reports go, what format an invoice uses or what style of reply the user expects. Memory should not, however, mean saving everything. When an agent stores thousands of pages of history and sends them to the model with every task, costs rise. The model has to process more tokens even when most of the data have no connection with the current matter.
OpenClaw can organize memory during periods of inactivity. This mechanism is sometimes called “dreaming”. The system reviews logs, extracts important facts and removes unnecessary details. It can be compared to an employee who, at the end of the day, rewrites chaotic notes into a short document. Instead of storing the entire conversation about selecting a date, the agent records only the result: “Meetings with customer X are best planned on Tuesdays after 12:00”. During the next task, it does not have to analyse dozens of earlier messages.
Well-designed memory is short, specific and capable of being corrected. Poorly designed memory is a collection of random sentences. The agent may then draw incorrect conclusions. If the user once asks for a formal email, the agent should not conclude that every future message is to be formal. Memory must distinguish a permanent rule from a one-off situation.
Where to run an agent
An agent can run on an ordinary computer, on a separate device or on a VPS server. Each option has specific consequences.
Installation on one’s own laptop is the cheapest because no hosting fee is paid. The problem appears when the lid is closed. The agent stops working. It will not check email at night, perform a task in the morning or answer a message from Telegram.
The second problem concerns security. An agent running on the main computer may receive access to private documents, photographs, passwords saved in the browser or devices on the home network. An agent error or a vulnerability in an installed add-on can then cause serious damage.
A separate old laptop, Raspberry Pi or small computer solves part of the problem. The agent can operate without access to the user’s main computer. It still depends, however, on electricity, home internet and self-administration. After a router failure or system update, it may stop working.
A VPS is a virtual server rented in a data centre. It runs around the clock and is separate from the user’s computer. The agent can be run in a Docker container, meaning a restricted environment. When something goes wrong, the effects are easier to contain within one server. For a beginner, a VPS is often a sensible compromise. It costs little, is available constantly and separates the experiment from the private computer. This does not remove the need for updates, backups and protection of API keys.
How much an AI agent really costs
OpenClaw and Hermes are open-source projects. Downloading the program itself does not require a licence fee. This does not mean, however, that the agent is free. Cost arises primarily in two places: on the server and through use of the language model. The simplest VPS usually costs several dollars per month. Lightweight tasks do not require a powerful machine. An agent that reads email, runs simple scripts and uses a model through an API can operate on a small server. In a practical budget, approximately $5–10 per month should be assumed for infrastructure. It is possible to go lower during a promotion, but a calculation should not be built on a price that applies only for the first month.
The second part of the cost is tokens. A token is a fragment of text processed by the model. The longer the instructions, the larger the memory and the more numerous the answers, the higher the bill. A cheap model can cost very little with simple tasks. A premium model used all day can cost many times more than the server.
A lightly used agent can stay within several dollars per month for models. An example is a daily report from five pages, a short summary of email and several conversations on Telegram. With this workload, the combined cost of the server and model may be approximately $8–20 per month. An intensively used agent costs more. When it checks a mailbox every few minutes, analyses long documents, uses an expensive model and maintains extensive memory, the bill can quickly rise to tens or hundreds of dollars. There is no upper limit. The agent can call the model hundreds of times a day.
OpenClaw does not charge a licence fee, but its extensive ecosystem encourages the installation of many functions. Each one can make additional requests. Frequent checking of system status, called a heartbeat, is particularly expensive. If the agent asks the model every five minutes whether anything important has happened, it performs 288 checks per day. Even a cheap individual request becomes expensive when multiplied by an entire month.
Hermes also does not charge for the program itself. Its cost advantage appears with repetitive tasks. Curator can turn a process based on many model calls into a simple script. The first runs may cost a similar amount to OpenClaw, but routine work may later be cheaper.
A real example looks like this. A small company runs OpenClaw on a VPS costing $7 per month. The agent analyses email once a day, creates a report once a week and responds to instructions several times a day. A cheap model generates a bill of $5. The whole thing costs approximately $12 per month. When the company switches all tasks to a premium model and adds a mailbox check every five minutes, the same OpenClaw can cost tens of dollars per month. With a very large number of documents, the cost can exceed $100. In the case of Hermes, the initial cost may look similar. The company pays an average of $7 per month for the VPS and several dollars for the model. The difference appears later. If Hermes saves weekly file processing as a ready-made script, the language model does not have to analyse the entire procedure every time. The bill for the repetitive task can therefore fall.
This does not mean that Hermes will always be cheaper. If the user commissions different, complicated analyses every day, Curator will not have many opportunities to use saved, repeatable procedures. In this case, cost depends primarily on the selected model and the length of the materials processed.
Materials concerning OpenClaw also recommend an initial top-up of the model provider’s account with approximately $40 or $50. This is not an OpenClaw subscription. These are funds intended for future API usage and obtaining higher limits.
The most important cost principle is simple. An expensive model should perform only tasks requiring difficult reasoning. A cheaper model or an ordinary script is sufficient for sorting messages, renaming files and simple reports. Otherwise, the company uses an expensive expert for work that a calculator can perform.
A schedule or continuous checking?
An agent can operate according to a fixed schedule or continuously check whether something new has appeared. A fixed schedule is cheaper and more predictable. The instruction “check the mailbox at 8:00, 12:00 and 16:00” launches three tasks a day. Continuous checking gives a faster response but is easy to configure badly. The instruction “check every minute” means 1,440 runs per day. If every run engages the language model, the bill rises even when nothing has happened.
A beginner should start with schedules. A morning report does not have to be checked every minute. Invoices can be checked once an hour. The status of a website can be checked by a simple script, and the language model should be run only when the script detects an error. A good solution separates detection of an event from its analysis. A simple and cheap mechanism checks whether a new file has appeared. Only after finding a file does it run the AI model. Thanks to this, the agent does not spend money on thousands of “nothing new” responses.
Security is not an add-on
An agent with access to email, a calendar and files has real power. It can send a message to a customer, delete a document or disclose confidential data. Security therefore cannot be the final stage of a project.
First, permissions must be restricted. An agent for organizing invoices does not need access to the whole of Google Drive. One folder is sufficient. An agent preparing replies does not need the right to send them automatically. It can create drafts.
The second principle is isolation. An agent running on a separate VPS should not have direct access to a private laptop. If it is attacked or carries out an incorrect instruction, the scope of the damage will be smaller.
The third principle concerns API keys. A key is the equivalent of a password to a service. It should not be placed in a message, a public file or code sent to a repository. It should be stored in protected environment variables on the server.
The fourth principle concerns add-ons. OpenClaw provides access to an enormous catalogue of skills, but not every plugin is trustworthy. The larger the ecosystem, the larger the attack surface. Installing a random add-on for an agent is more serious than installing a simple word-processor plugin because the agent may have broad permissions. Hermes has a smaller ecosystem of external add-ons, which reduces part of the risk. On the other hand, its mechanism for automatically creating skills may be less predictable for a beginner. The system may optimize a procedure itself. This is convenient but requires logs, change control and the ability to reverse an incorrect version.
The fifth principle is the most important: at the beginning, do not connect a bank account, password manager or system whose failure could cause a large loss. The first agent should work on test data, copies of documents and restricted accounts.
It is also worth requiring confirmation of destructive actions. The agent may propose deleting a file, but it should not perform this operation independently. It may prepare a transfer for checking, but it should not approve it. It may find a flight, but it should not buy a ticket without presenting the price and obtaining consent.
Simple tasks to begin with
A good first project should not make important decisions independently. It can organize files, prepare draft replies, summarize reports or keep track of deadlines.
An example is an invoice agent. Once a day it checks the specified folder, reads supplier names and dates, renames files according to an agreed scheme and prepares a table. It does not make payments. It does not delete documents. When it is unsure of a result, it moves the file to a “to be checked” folder.
Another example is a recruitment agent. It reads new messages from recruiters, extracts the job title, rate, working arrangement and requirements, and then saves the data in a spreadsheet. It can prepare a reply but does not send it without consent.
A home agent can check the weather, calendar and journey time every morning. It can write: “You have a meeting in the city centre at 14:00. The expected journey will take 45 minutes. Rain is possible in the afternoon”. This is a simple task, but it demonstrates all the characteristics of an agent: memory, a schedule, tools and active notification.
An agent in an online store can analyse customer questions once a day. If a question concerns shipment status, it retrieves data from the logistics system and prepares a reply. If a customer makes a complaint, the agent does not make a decision. It passes the matter to an employee together with a short summary.
An agent for a salesperson can prepare a list each morning of customers who should be contacted. It checks the date of the last conversation, open offers and purchase history. It does not call by itself. It provides the salesperson with ready-made material for work.
Which system should you choose?
OpenClaw is a better starting point for someone who wants to see a working panel quickly, connect popular services and use ready-made skills. Add-ons must, however, be chosen carefully and the system updated regularly.
Hermes suits a user who wants to automate their own repeatable processes and accepts a more technical mode of operation. Its Curator may reduce costs over time because it turns part of the work into ordinary procedures.
A person running a small company who needs a quick connection to email, a calendar and Google Drive can start with OpenClaw. A person who analyses similar files every week and generates reports according to their own scheme may gain more from Hermes. For a complete beginner, the choice of the first task is more important than the choice of name. A badly chosen project will be costly and dangerous regardless of the platform. A well-chosen one will make it possible to understand how an agent works without great risk.
The most sensible start is a separate VPS, one test account, one messenger and one simple task run at a fixed time. The agent should prepare the result and a human should approve actions. Permissions can be increased only after several weeks of stable work.
An agent does not replace responsibility
The greatest mistake is to treat an agent as an infallible employee. A language model can misunderstand an instruction, confuse data or invent missing information. In addition, an agent can carry out this error in a real system. If a chatbot writes an incorrect reply, the user can correct it before sending. If an agent has the right to send messages independently, the error reaches the customer immediately. If it has the right to delete data, an incorrect decision can lead to the loss of documents.
A good agent therefore does not operate without boundaries. It has a clearly described goal, restricted access, an action log and approval points. It can work independently on simple operations. For financial, legal or irreversible actions, it should stop before execution and ask a human for a decision. This does not reduce the value of agents. On the contrary. Well-set boundaries mean that they can be entrusted with more routine work. A human does not have to copy data, review the same folders and write similar replies. They nevertheless retain control over what is important.
An agent on a private VPS server
Installing an agent on a private VPS server is the recommended solution for most users, especially beginners. Sources indicate that this is the most stable way to have a personal AI assistant.
Here is a detailed summary of the purpose of such an installation and its advantages.
The main purpose is to create an autonomous assistant that works independently of personal equipment. By placing the agent in the cloud, it can monitor the network, manage email or perform scheduled tasks (cron) around the clock, seven days a week, even when your computer is switched off. This allows complete mobility—you can communicate with the agent using a telephone (for example, through Telegram) from anywhere in the world.
Security and isolation (sandbox): a VPS creates an isolated environment. If something goes wrong or the agent is attacked, only the virtual machine is threatened, not your private files on your home laptop.
Easy management: in the event of configuration errors, you can easily restore the system from a backup or wipe the server and deploy it again in a few minutes.
What are the disadvantages of this solution?
Unlike installation on one’s own old laptop, a VPS involves a monthly fee, usually from several to more than ten dollars.
Complexity of initial configuration: despite the easy installation of the system itself, you must still obtain API keys from providers such as Anthropic and OpenAI yourself and configure tokens for messengers.
Dependence on a provider: you use the infrastructure of a third-party company, which entails the need to accept its terms and privacy policy.
Additional costs: in addition to the server fee, you must pay AI-model providers for every token (word) processed, which, with badly configured automated tasks, may generate unforeseen expenses.
What does an AI agent really change?
An AI agent changes the way artificial intelligence is used. A chatbot such as ChatGPT is a tool for conversation. An agent becomes part of a process. It remembers rules, uses services, launches tasks and returns with a result.
OpenClaw demonstrates the power of a large ecosystem of ready-made skills. Hermes demonstrates the power of automatically learning repetitive procedures. The first provides faster access to many integrations. The second may fit a specific way of working better and reduce the cost of routine tasks. Both systems can be run without a licence fee, but both generate bills for the server and AI models. With sensible use, a beginner can stay below $20 per month. With bad configuration, the same agent can consume many times more.
The most important question is therefore not whether an agent can act by itself. It can. The most important question is: exactly which actions should it perform, with which permissions and with what cost limit?
A good agent is not a digital, omniscient helper. It is a precisely configured executor. It has clearly defined work, access only to the tools it needs and an obligation to stop where risk begins. In this form, an AI agent ceases to be a technology demonstration. It becomes a useful component of everyday work.
Wojciech Moszczyński
Wojciech Moszczyński—a graduate of the Department of Econometrics and Statistics at Nicolaus Copernicus University in Toruń; a specialist in econometrics, finance, data science and management accounting. He specializes in the optimization of production and logistics processes. He conducts research in the area of the development and application of artificial intelligence. For years, he has been involved in popularizing machine learning and data science in business environments.
Dodaj komentarz