The Concept of Agency in Artificial Intelligence
What does agency in artificial intelligence really mean? Explore the concept of agency, its core ingredients, and how it separates agents from tools.
Agency is the quality that distinguishes a system that merely responds from one that acts on its own behalf toward a goal. In artificial intelligence, agency describes a system's capacity to perceive a situation, decide what to do, and take action to influence outcomes. It is the conceptual foundation beneath the entire field of agentic AI, and understanding it clarifies why some systems are called agents while others are simply tools.
Defining Agency
In its broadest sense, an agent is anything that perceives its environment through sensors and acts upon that environment through effectors. This classical definition, drawn from the study of intelligent systems, frames agency around a perception-action loop. A thermostat technically fits this definition: it senses temperature and acts by switching heating on or off. But meaningful agency in modern AI implies more than a reflexive response; it implies the system selects among possible actions in pursuit of an objective, rather than executing a single fixed behavior.
The richer notion of agency adds the idea of goal-directedness. An agent does not just react; it acts in order to bring about a preferred state of the world. That orientation toward goals is what separates a passive function from an active participant in a task.
The Ingredients of Agency
Several capabilities combine to produce genuine agency. The first is perception, the ability to take in information about the current state of the environment. The second is a representation of goals or preferences, some internal notion of what outcomes are desirable. The third is the capacity to reason or deliberate, weighing options against those goals. The fourth is the ability to act, to do something that changes the environment or moves toward the goal. Underlying all of these is autonomy, the degree to which the system makes these choices itself rather than having every step dictated externally.
When these ingredients are present, a system can be said to exercise agency. It senses where things stand, considers what it wants to achieve, chooses an action it judges likely to help, and carries it out. The loop then repeats with updated perception, allowing the agent to respond to the results of its own actions.
Degrees of Agency
Agency is not binary; it exists along a spectrum. At the low end are systems with narrow, tightly constrained autonomy, acting only within rigid limits and deferring most decisions to humans or fixed rules. At the high end are systems that set sub-goals, choose their own methods, and operate for extended periods with little intervention. Most practical AI agents sit somewhere in between, granted enough autonomy to be useful while remaining bounded by the permissions and oversight their designers impose.
This graduated view matters because it frames agency as a design choice rather than an all-or-nothing property. How much agency to grant a system is a deliberate decision, balancing the usefulness of independent action against the risks of letting a system act without supervision. More agency can mean more capability, but it also means more responsibility placed on the system and more careful guardrails required around it.
Why Agency Matters in Practice
Treating agency as a real property changes how teams design and govern AI. A tool is judged on the quality of its output; an agent must be judged on its behavior over time, including the actions it chooses, the consequences it causes, and its alignment with the intent behind its goals. Because an agent can act, it can also act wrongly in ways that are harder to contain than a flawed piece of generated content. This is why discussions of agency are inseparable from discussions of safety, alignment, and control.
Recognizing where agency begins also helps set expectations. A system without genuine agency will not independently pursue objectives no matter how capable it seems, while a system with real agency requires the kind of oversight appropriate to something that makes its own decisions. Naming the concept clearly keeps both designers and users honest about what a given system is actually doing.
Frequently Asked Questions
Is every AI system an agent?
No. Many AI systems, such as classifiers or pure text generators, produce outputs without perceiving an environment or pursuing goals through action. Agency specifically requires the capacity to sense, decide, and act toward an objective, which only some systems possess.
Does more agency always make an AI better?
Not necessarily. Greater agency increases capability but also increases the risk of unintended actions and the need for oversight. The appropriate level of agency depends on the task, with high-stakes or irreversible actions usually calling for tighter constraints.
How is agency different from intelligence?
Intelligence concerns how well a system reasons or solves problems, while agency concerns whether it can act on its own behalf toward goals. A system can be highly intelligent yet have little agency, or have meaningful agency while being fairly limited in raw capability.