agent
in English: AI agent
An agent is an artificial intelligence program that not only responds but independently plans, uses tools, and completes tasks in multiple steps to achieve a given goal.
In plain terms
A chatbot is like a knowledgeable friend: you ask, it answers, and you do the rest. An agent is more like an assistant to whom you say: 'organize Thursday's meeting,' and it checks calendars, finds a time, sends invitations, and then tells you it is done.
The essence is therefore action. The agent receives a goal, not a precise list of instructions. It decides itself what steps are needed and uses tools for this: it searches the web, opens files, runs programs, fills out forms, and writes emails.
Meanwhile, it also monitors what has happened. If one step fails, it tries something else, just as you would choose another route if the street is blocked. This makes it different from traditional automation, which always executes the same pre-written sequence of steps.
It is good for taking over longer, multi-step, boring, or time-consuming tasks from you. The price of autonomy is that you must pay closer attention: what access you grant it and what you check afterward.
How it works
- It receives a goal: you write in natural language what you want to achieve and specify which tools and data it can access.
- Planning: the agent's 'brain' is typically a large language model that breaks the goal into smaller steps and decides where to start.
- Acts: it invokes a tool, such as a search engine, calendar, database, browser, or code execution environment, and receives its result.
- It observes and corrects: it checks what the step produced and decides the next action based on that. This planning–action–observation loop repeats until the task is complete.
- Stops and reports: when finished or stuck, it summarizes what it did. A well-configured agent also stops before risky steps (such as payment, deletion, sending) and requests approval.
An everyday example
You tell your phone assistant: 'Find accommodation in Eger for the weekend for four people under 60,000 forints, and book the best one.' The agent browses multiple pages, compares offers, fills out the booking form, and asks for confirmation before payment.
Who is affected
A common misconception
Limits and risks
- It can make mistakes and be confidently wrong; in many interdependent steps, small errors accumulate, so always verify important results.
- Grant it only the access the task requires, because whatever it can access (email inbox, bank card, files), it can also misuse.
- Deceptive: a hidden instruction in a website or email can compel you to do something you did not request (this is called prompt injection).
- Responsibility remains with you or your company, and running long, multi-step tasks can be slow and expensive.
Similar, but not the same
- chatbot – A chatbot answers questions and converses, while an agent acts independently in multiple steps to achieve your goal.
- nagy nyelvi modell (LLM) – The language model is only the agent's 'brain'. It becomes an agent when it is given tools and an independent work cycle, often including memory.
- traditional automation (RPA, scripts) – Automation repeats a pre-written, rigid sequence of steps, whereas the agent itself decides the steps and adapts to unexpected situations.
Related concepts
coding agent, multi-agent system, browser agent, Customer service agent, research agent
large language model, chatbot, tool use, prompt injection, human oversight
A brief history
The concept of the autonomous sensing and acting 'intelligent agent' has been a fundamental term in AI research for decades. It was widely popularized by Stuart Russell and Peter Norvig's 1995 textbook (Artificial Intelligence: A Modern Approach). Large language model-based agents came into the spotlight in 2023, with early experimental projects (such as AutoGPT) and the models' tool-calling capabilities. More reliable, practically usable versions spread in 2024–2025.
Where things stand
At the end of 2026, coding and research agents are already everyday tools in many places and have been built into the major office software suites. However, fully autonomous, unsupervised operation is still rare: in most places a human approves the important steps.
In the news
- Ethan Mollick publishes Portals, a browser-based 3D experience of 471 generated worlds built by agent teamsNov 9, 2026 · Code
- AI coding agents still need human rework on three-quarters of tasksNov 8, 2026 · Code
- Attackers use agentic tooling to find exposed secrets in public GitHub reposNov 6, 2026 · Code
- GitHub Copilot impact dashboard adds chat and VS Code agent usage, code review gets API controlsNov 6, 2026 · Code
- Google Cloud adds OpenAI models and Codex agents to Model GardenNov 6, 2026 · Code
- GitLab 19.5 makes agentic development a governed part of its DevOps platformNov 6, 2026 · Code
- JP Morgan Payments and Mastercard invest in agentic commerce startup firmlyNov 6, 2026 · Money
- Almost half of Visa's AI shopping agent test runs failNov 5, 2026 · Marketing
Also known as
agent · AI-agent · AI agent · MI-agent · agent · intelligens agent · intelligent agent · autonomous agent
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.