agent-ciklus

in English: agent loop

advancedwidespread

The agent cycle is the repeating loop in which an AI agent thinks, acts, reviews the result, and decides again until the task is completed.

In plain terms

A plain chatbot answers your question once and is done. An agent, by contrast, works in cycles: it takes a step, sees what resulted, and decides the next move based on that. This cycle is called the agent loop.

It is like when you are making soup. You do not throw everything in at once: you taste it, add salt, taste it again, and repeat until it is good. The agent proceeds similarly, except that instead of tasting, it might run a search, open a file, or try a program.

Due to the cycle, the agent can independently complete multi-step tasks and correct its own mistakes along the way. However, this makes it slower and more expensive than a simple response, as each round requires additional computation.

How it works

It receives the goal: you write what you want, and the agent stores it in its working memory, the context.
Decides on the next step: the language model assesses what is still missing and selects a tool, such as search, file reading, or code execution.
Acts: the program running around it actually executes the requested operation, because the model itself only writes text and does not touch the outside world.
It observes the result: the operation's output, even an error message, is fed back into the context, so the next decision is made with this knowledge.
It repeats or stops: the loop continues until the agent deems the task complete or reaches a set limit, such as the allowed number of steps, time, or cost. Before a sensitive step, the cycle may pause to await human approval.

An everyday example

You ask an AI assistant to find weekend accommodation in Eger for two adults, with parking. You watch it work step by step: searching, opening a few pages, comparing prices, searching again because one place is full, and only at the end making a recommendation. This sequence of steps is the agent loop.

Who is affected

developerIn this cycle, the coding agent writes, runs, and fixes the program, so you must set when it should stop and what it can access. → Code

employeeIf you understand that the agent works in loops, you can better judge when to entrust a task to it and when to intervene. → Work & HR

entrepreneur, executiveEvery cycle costs money and time, so the cycle length directly determines how much it costs you to operate an agent.

everyoneMore and more everyday assistants work this way, and it is good to know that longer work does not automatically mean more reliable results.

A common misconception

MythThe agent learns during the loop and gets smarter with every round.

In realityThe model's knowledge does not change during this; only more information about the specific task enters its working memory. For the next task, it typically starts with a clean slate unless it was given separate memory.

Limits and risks

An early error can ripple through subsequent rounds because the agent builds its next step on its own flawed result.
An agent can get stuck and repeat the same failed step, so a step, time, or cost limit is always required.
In long cycles, the working memory can fill up. In such cases, previous steps must be compressed or omitted, and the agent may lose the thread or forget your original request.
If the agent performs real operations, such as sending an email, deleting a file, or making a payment, a faulty loop has real consequences, so request approval for irreversible steps.

Similar, but not the same

chain of thought – A chain of thought is the model's step-by-step reasoning performed internally, whereas in an agent cycle, thinking is followed by actual action and its results.
munkafolyamat (workflow) – In a workflow, a human pre-writes the sequence of steps; in an agent loop, the model decides step by step what comes next.
training loop – The training cycle runs when the model is being trained and changes the model itself; the agent loop occurs while using the finished model.

Related concepts

Broader

AI-agent, agent-based AI, autonomous system

Narrower

tool calling, stopping condition, planning, self-check

On the edge

context window, agent-keretrendszer, human approval, multi-agent system, coding agent

A brief history

The perception–decision–action loop is an old concept in artificial intelligence research. The 2022 ReAct paper described the alternation of thinking, acting, and observing for language models. After that, the term 'agent loop' spread in its current meaning.

Where things stand

By the end of 2026, almost every AI agent is built on this cycle, from coding agents to research and browsing assistants. Current development focuses mainly on ensuring the cycle remains reliable, verifiable, and affordable even for long tasks.

Status: 2026 November

Also known as

agent loop · agentic loop · agent-hurok · agentciklus · agent loop · ReAct-ciklus · think-act-observe cycle · think-act-observe loop

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.