autonomy

basicspreading

Autonomy shows how much an AI agent can decide and execute independently without seeking human approval at every step.

In plain terms

Imagine a new colleague. During the first week, you review every email before they send it. After a month, they only check in on the more important matters. After six months, they run the project alone and only show you the final result. Autonomy is the same scale for AI agents: how much room to act the machine is given.

Autonomy is not a switch but a spectrum. On one end stands the chatbot, which only answers and requires you to execute everything. On the other end stands the agent, which receives a goal, then plans, uses tools, and corrects errors independently for hours, reporting only at the end.

It is useful because it saves time: you do not have to be present at every small step. The price is that the more autonomous the agent is, the later you notice if it has gone in the wrong direction. Therefore, autonomy should always be adjusted to the size of the stake.

How it works

You give a goal, not steps: you do not specify what to click on, but what you want to achieve.
The agent breaks the goal into subtasks and decides itself in what order and with what tool (search, file management, email, program execution) to solve them.
After each step, it reviews the result and decides on the next action accordingly. If something fails, it tries something else.
Permissions define its scope of action: it can be configured to determine what it can do freely, what requires your approval (for example, payment, deletion, sending an email), and what it can never do.
At the end, it reports back, or it stops if it gets stuck, reaches its limit, or comes to a decision that a human must make.

An everyday example

You ask your AI assistant to organize a three-day trip to Vienna. With low autonomy, it only shows offers, and you book. With high autonomy, it selects the train and accommodation itself, books them, and asks for confirmation at most before payment.

Who is affected

employeeYou can delegate more routine tasks to an agent, but you must decide what to let go of and what to verify. → Work & HR

entrepreneur, executiveYou must define what the agent may decide on its own at the company, and who is responsible if it makes a mistake.

developerYou set the autonomy level of the coding agent: whether it should ask before every command or write, test, and fix independently. → Code

lawyer, citizenThe EU AI Act explicitly mentions autonomy in its definition of an AI system and requires human oversight for high-risk systems. → Law

A common misconception

MythAn autonomous agent has its own will and does what it wants.

In realityAutonomy here means only that the agent independently chooses steps within the given goal and permissions. The human sets the goal and boundaries, but the agent can still make mistakes or misunderstand the task within these limits.

Limits and risks

Errors accumulate: in a long, independent session, every subsequent step may be built on an early mistake.
Greater autonomy requires greater access (accounts, files, a bank card), so a mistake or a malicious text that deceives the agent can cause more damage.
Responsibility does not transfer to the machine: you or your company are typically liable for what the agent does on your behalf.
For irreversible actions (payments, deletions, public publication), it is advisable to retain human approval.

Similar, but not the same

automation – Traditional automation works according to pre-set rules and steps, while an autonomous agent chooses its own steps by adapting to the situation.
autonomous driving – Autonomous driving refers to vehicle autonomy with its own level system ranging from 0 to 5; agent autonomy concerns the independence of software in task execution.
artificial general intelligence – AGI means a system performs at least at human level across a very wide range of tasks; autonomy refers to how much it can do independently. An agent capable of a narrow task can still be highly autonomous.

Related concepts

Broader

AI-agent, agentic AI, artificial intelligence

Narrower

autonomy levels, fully autonomous agent, semi-autonomous operation

On the edge

human oversight, ember a folyamatban, permission management, safeguards, tool use

A brief history

The word is of Greek origin (autos = self, nomos = law), originally meaning self-government or self-determination according to one's own laws. As a fundamental property of agents, it was established in Michael Wooldridge and Nicholas Jennings' 1995 study, "Intelligent Agents: Theory and Practice." The EU's 2024 AI Act also incorporated "different levels of autonomy" into the legal definition of an AI system.

Where things stand

At the end of 2026, some coding and office agents can already work independently on a single task for hours, but most companies retain human approval for high-stakes steps. Fully unsupervised operation is rare.

Status: 2026 November

In the news

Appears in 6 Agent Age stories; the curve shows weekly occurrence in the news magazine.

UK regulators say existing laws apply to AI agentsNov 2, 2026 · Jog
Caterpillar and FieldAI target AI-driven quarry operationsSep 14, 2026 · Ipar
Visa CEO: AI shopping has arrived, but agentic payments haven'tSep 9, 2026 · Money
Caterpillar and FieldAI to develop physical AI for construction and industrial sitesSep 7, 2026 · Ipar
Agricultural robotics and AI: from China's grain heartland to Carbon Robotics and Bonsai RoboticsAug 31, 2026 · Ipar
Carmakers' autonomy race shifts to AI-powered Level 2 systems over eyes-off Level 3Aug 25, 2026 · Ipar

Also known as

autonomy · autonomy · independence · autonomous agent · autonomous agent · autonomous agent · Autonomy level · level of autonomy

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