agentikus munkafolyamat

in English: agentic workflow

advancedspreading

A working method where AI does not give a single answer but plans in multiple steps, uses tools, checks the result, and corrects it.

In plain terms

A simple chatbot responds in a single attempt. This is like asking someone to write an essay from start to finish without deletion or rereading. In an agentic workflow, the AI works like a careful person: it creates an outline, researches, writes, rereads, and corrects. Today's chat assistants often work this way in the background, for example when searching the web.

It is similar to renovating an apartment. It is not one master doing everything in one afternoon, but there is a plan, a sequence, several specialists, and a final handover. The workflow here also consists of steps. Some steps are performed by an AI model or an agent, others by traditional software, often with external tools: a search engine, a spreadsheet, email.

It is useful for enabling AI to perform complex, multi-step tasks more reliably, not just answer questions. Traditional automation follows rigid rules. Here, the AI itself decides at certain points what the next step should be.

How it works

You specify the goal, for example, to prepare a summary of competitors' prices. The system breaks this down into subtasks.
For individual steps, the AI calls on tools: it searches the web, opens files, queries a database or runs a program.
The result of each step is input for the next. The system keeps track of where it is and what it has learned.
A control step reviews the partial result, which can be performed by another agent. If an error is found, the system steps back and corrects it.
It can pause at key points and request your approval before, for example, sending an email or initiating a payment.

An everyday example

You upload an invoice to the accounting software. The AI reads the data, compares it with the order, checks the tax number, records the entry, and only notifies you if something is wrong. You see a single button press, while five or six steps ran sequentially in the background.

Who is affected

employeeSuch processes take over some of your repetitive, multi-step office tasks, leaving you increasingly with oversight and decision-making. → Work & HR

entrepreneur, executiveYou must decide which processes are worth automating this way and where human approval must remain mandatory.

developerYou design the steps, tools and checkpoints, and a coding agent works in a similar workflow in your own work too. → Code

marketerResearch, copywriting, review, and scheduling can be linked into a single chain, making campaign preparation much faster, but you must still review the result. → Marketing

A common misconception

MythAgentic workflow means that AI works completely independently, without human supervision.

In realityMost such processes consist of pre-planned steps, and the AI only makes free decisions at certain points. In well-built systems, a human approves continuation before risky steps.

Limits and risks

Errors accumulate: if an early step is wrong, later steps build on that mistake, and the final result can be confidently wrong.
More steps mean more model calls, making it slower and more expensive than a single question-and-answer.
If the AI has access to tools and data, the damage caused by an error or a malicious instruction hidden in the text can be greater.
It is harder to determine later at which step and why something went wrong, so do not entrust important tasks to it without logging and human oversight.

Similar, but not the same

AI-agent – The agent is the actor that independently decides and acts, while the agentic workflow is the framework in which one or more agents perform the task step by step.
traditional automation (RPA) – Traditional automation always follows the same rigid rule set, whereas in an agent-based workflow, the AI adapts the steps to the situation.
prompt chain – A prompt chain executes instructions in a fixed order, while an agentic workflow additionally uses tools, verifies, and steps back if necessary.

Related concepts

Broader

agentic AI, automation, artificial intelligence

Narrower

prompt chain, multi-agent system, self-check (reflection), Task planning

On the edge

AI-agent, tool use, orchestration, human oversight, coding agent

A brief history

The term was popularized in 2024 by Andrew Ng, founder of DeepLearning.AI and instructor at Stanford University: in his lectures, he showed that models working in multiple steps with self-correction yield better results than those responding in a single attempt. At the end of the same year, Anthropic's paper titled "Building effective agents" distinguished between pre-scripted workflows and agents that direct their own steps.

Where things stand

At the end of 2026, many large office, customer service and developer software products offer built-in agentic workflows, and in software development they are already widely used. At companies, many deployments are still at the experimental stage, mainly because of questions of reliability, cost and liability.

Status: 2026 November

In the news

Appears in 3 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Self Labs launches Agent Vault as banks lock AI agents out of accountsOct 16, 2026 · Money
Working group releases 'Skills over MCP' specification for agentic workflowsSep 15, 2026 · Code
California AI disclosure rules become operative as NAI issues adtech guidanceAug 3, 2026 · Marketing

Also known as

agentic workflow · agentic workflows · agentikus workflow · agent-based workflow · agent-based workflow · agentic workflow · AI agent workflow

Sources

Anthropic: Building effective agents

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