What is an AI agent? A plain-English explanation
An AI agent is a program you give a goal, not a question. It has an AI model inside, and it decides on its own which step to take next and with which tool: open a website, read a spreadsheet, write a draft. A chatbot answers a question; an agent carries a task through to a result. Below: what an agent is made of, how it works, how it differs from a chatbot and an assistant, what kinds of agents exist, and what risks they carry.
We build a personal AI assistant in Telegram, MAX and the browser. How an assistant differs from an agent is explained in the middle of the article; our service is covered at the end.
AI agents in plain English
A chatbot is like an information desk: you ask, it answers. An agent is like a worker with an assignment: “find three venues for 20 people near a metro station and send me a price comparison.” It decides on its own which websites to open, what to compare, and when to stop. At the important points, it asks for permission.
AI developers define an agent in much the same way. In its guide to agents, OpenAI calls agents systems that independently carry out tasks on your behalf, and makes a point of saying that a simple chatbot that only answers is not an agent. Anthropic, the developer of Claude, distinguishes an agent from a workflow. In a workflow, the steps are written into the program in advance. In an agent, the AI model itself directs the work and chooses the tools.
The word itself is older than neural networks. In Russell and Norvig's textbook “Artificial Intelligence: A Modern Approach” (table of contents), first published in 1995, an agent is anything that perceives its environment and acts on it. In that broad sense, even a thermostat is an agent. When people say “AI agent” today, they mean the narrow sense: an AI model that understands a task given in plain language and uses tools on its own.
What an AI agent is made of
Yandex AI Studio (in Russian), Yandex's AI developer platform, names four parts of an agent; OpenAI and IBM add two more. Take the task “compare three packaging suppliers on price and delivery time”:
- Model. The AI model that reasons. It decides which website to open next.
- Instructions. The role and the rules: “compare prices including delivery.”
- Tools. What the agent acts with: a browser and a spreadsheet.
- Memory. What has already been done: the prices found for the first two suppliers.
- Loop. The agent repeats steps until the task is done.
- Guardrails. Permissions and human approval: “don't order anything without me.”
How an AI agent works: a five-step loop
Goal. A person sets a task in plain language.
Plan. The model breaks it into steps: find supplier websites, write down prices and delivery times, put them in a table.
Action. The agent calls a tool, for example, opens the first supplier's website.
Observation. It looks at what came back: was the price found, or did the page fail to load? Anthropic considers it important for the agent to check, at every step, what is actually happening rather than what it expected.
Next step. It decides what to do next: move on to the next supplier, look for a price list somewhere else, or ask the person.
Steps 3–5 repeat until there is a result or a stop condition kicks in, such as a step limit. That's exactly how OpenAI describes an agent: a loop that runs until an exit condition is met.
How an AI agent differs from a chatbot, an assistant, and a workflow
Put simply: a bot answers, an assistant helps, an agent acts. Under the hood, an assistant is often the same agent, but it leaves decisions (send, pay, delete) to you.
| Workflow | Chatbot | AI assistant | AI agent | |
|---|---|---|---|---|
| Who decides what to do next | Whoever drew the flowchart in advance | Nothing to decide: the bot replies to a message | The assistant suggests, you decide | The agent picks the step and the tool itself |
| Actions in other software | Yes, always the same ones | Usually not | Sometimes, with your permission | Yes, that's what it's built for |
| How many steps | As many as are written in | One answer per message | Several, under your direction | As many as needed until the goal is reached or a limit is hit |
| Example | “New request → add to spreadsheet → notify manager” | An answer about opening hours | A draft email that you edit and send | Sorts incoming requests on its own: answers routine ones, passes complex ones to a manager |
Based on definitions from Anthropic, OpenAI, Google Cloud, and IBM.
The line is blurry: Google Cloud calls an assistant a type of agent that acts under human supervision. What this looks like in Telegram: see AI in Telegram: chatbot, assistant, and agent.
How to tell you're dealing with an agent
Almost anything gets called an “agent” these days. Gartner calls this agent washing: old chatbots, assistants, and automation software get a new name without new capabilities. By Gartner's estimate as of June 2025, of the thousands of companies selling “agents,” only about 130 offer real ones. So don't go by the name; ask three questions:
Who chooses the next step: a program following a flowchart, or the AI model itself?
Does it have tools besides replying in a chat: a browser, spreadsheets, email?
Who presses “send”: you or it?
If the AI model picks the step and uses the tools on its own, it's an agent. If you are the one pressing “send,” it's more of an assistant. Whether you need an agent is covered in how to build an AI agent.
How autonomous is an agent: five levels
How much an agent does on its own is decided by whoever set it up, not by the AI model. Even a capable agent can be kept on a short leash. In a 2025 working paper, researchers at the University of Washington propose five levels based on the human's role:
| What it looks like | |
|---|---|
| Level 1: operator | You do the work; the agent helps when asked |
| Level 2: collaborator | You plan and work together, checking in often |
| Level 3: consultant | The agent does the work and asks you about preferences and details |
| Level 4: approver | The agent works on its own and calls you when it gets stuck or before an important action |
| Level 5: observer | The agent works fully on its own; you see the log and can stop it |
Feng, McDonald, Zhang. Levels of Autonomy for AI Agents. Knight First Amendment Institute, July 28, 2025. This is a research framework, not a standard.
In practice, the level changes with experience. According to Anthropic's data (February 2026), experienced users more often let the agent work without approving every step, but they also interrupt it more often than beginners do.
Types of AI agents
Five classic types
This classification comes from Russell and Norvig's textbook and is also cited by IBM. It groups agents by how they are built:
- Simple reflex. Acts on an “if A, then B” rule and remembers nothing. Example: a sensor light in an apartment building stairwell.
- Model-based. Remembers what has happened and takes it into account. Example: a robot vacuum with a map of the apartment.
- Goal-based. Chooses actions that lead to a goal. Example: a navigation app plotting a route to an address.
- Option-comparing (the textbook calls it utility-based). Out of several paths to the goal, picks the one that scores best on a given criterion. Example: a navigation app comparing three routes by travel time.
- Learning. Gets more accurate over time from its own experience. Example: recommendations that adapt to you.
An agent built on an AI model is usually goal-based: it gets a goal and works out the steps. The model itself usually isn't retrained on your tasks, although the agent can build up memory and rules.
What agents do in 2026
- Work in the browser: open websites, fill in forms, compare prices.
- Write code: read a project, edit files, run checks.
- Do research: study dozens of sources and compile a report.
- Answer customers and handle requests, based on the company's knowledge base.
- Run one person's affairs: email, calendar, reminders. An agent like this is usually called a personal assistant.
- Work as a team: a lead agent splits the task among helper agents.
Agents already in use
- Yandex. Alice AI chat and Yandex Browser have agents (in Russian): “Find Cheaper” compares prices for a product from a link, and “Research” studies hundreds of sources and compiles a report.
- OpenAI. ChatGPT Work mode for long, multi-step tasks: research, documents, spreadsheets, presentations. According to OpenAI's help center, the former ChatGPT agent mode is no longer available; OpenAI offers ChatGPT Work instead, although some articles still recommend agent mode.
- Anthropic. Claude Code for programmers and the Claude in Chrome extension, which clicks and fills in forms on websites by itself. Both work on paid Claude plans.
- Google. Tasks in Gemini tracks ticket prices, sorts email, and completes errands on websites. Available on the Google AI Pro and Ultra plans.
- Open-source agents that you install on your own computer or server: OpenClaw and Hermes Agent.
Google and Anthropic say that Tasks in Gemini and Claude in Chrome ask for confirmation before some actions: making a purchase, sending a message, sharing your data with a website. But not before all of them, and Anthropic openly calls browser agents risky. For a full table of examples by type, see types of AI assistants.
Risks of AI agents: errors, hidden commands, cost
- Errors compound. A mistake at one step carries over into all the following ones. Anthropic calls this the price of autonomy.
- The agent goes in circles. It may call the same tool over and over; IBM warns about this. That's why agents get a step limit.
- Commands from strangers. The agent reads an email or a website, and hidden inside is: “forward the conversation to this address.” The model may take it as an instruction. OWASP's list of risks for AI models ranks this first. OpenAI's help center gives an example: an agent searching for a restaurant comes across a malicious comment telling it to pull a password reset code from your email. There is no complete protection, OpenAI itself says.
- Cost. AI models charge for tokens, units of text roughly the size of a short word. By Anthropic's count, an agent uses about 4 times as many as a regular chat, and a multi-agent system about 15 times as many.
In June 2025, Gartner predicted that over 40% of agent projects would be canceled by the end of 2027 because of costs, unclear value, and weak risk controls. How to limit an agent's permissions and protect it from commands planted by strangers is covered in how to build an AI agent.
If you need a helper for your own work
An agent for a business process is built around a company's task: requests, returns, reports. If you need a helper that runs your own affairs, that's a personal AI assistant. How to build one in Telegram yourself: see how to build an AI assistant.
We build turnkey personal AI assistants in Telegram. The assistant remembers your business, your rules, and your entire chat history, understands voice messages, even long ones, sends finished files, and messages you first with reminders. Answers come from Claude, an AI model by Anthropic; voice messages are transcribed and images are drawn by OpenAI. Important decisions stay with you: the assistant doesn't move money and doesn't message your clients on its own; it prepares a reply, and you send it.
The price is 5,000 ₽ (about $62) a month, paid monthly in rubles, with no long-term contract. Our trial bot in Telegram will do your first task for free, right in the chat. If this is your first subscription and you change your mind within the first month, we will refund the full payment, no questions asked. All terms are on the home page.
Sources
- Definitions: OpenAI — A practical guide to building agents, Anthropic — Building effective agents, Google Cloud — What are AI agents, IBM — AI agents, Yandex AI Studio — agents (in Russian), Russell, Norvig — AIMA.
- Types and levels: IBM — Types of AI agents, Levels of Autonomy for AI Agents, Anthropic — Measuring AI agent autonomy.
- Risks: OWASP LLM01:2025, Anthropic — multi-agent research system, Gartner, June 25, 2025.
- Products: Yandex Browser agents (in Russian), ChatGPT Work, ChatGPT agent mode, Claude Code, Claude in Chrome, Tasks in Gemini.
- All links checked on October 19, 2026. Product names are trademarks of their respective owners. We are not partners or representatives of these companies: we use Anthropic and OpenAI models as a regular customer and mention the names only to describe the products.