An AI agent, also known as agentic AI, is an autonomous entity that perceives its environment through sensors and acts upon it through actuators to achieve specific goals. This concept is central to the agent-based definition of artificial intelligence, which describes AI as the study and design of intelligent agents. Agents can range from simple control systems, such as a thermostat, to highly complex systems like a self-driving car or a human being, and they may improve their performance over time through Machine learning or by acquiring knowledge.
The term "intelligent agent" originates from the broader field of Artificial intelligence, where textbooks by Stuart Russell and Peter Norvig define it as anything that perceives its environment and performs actions. A rational agent selects actions expected to maximize its performance measure, given its percept sequence, prior knowledge, and available actions. This rationality does not require omniscience or guaranteed success; it concerns the expected outcome based on available information. Agents are also characterized by properties such as autonomy, responsiveness to environmental changes, and goal-directed or proactive behavior, often modeled using architectures like belief-desire-intention (BDI).
Objective Function
The objective function, also called a goal function, specifies the goals of an intelligent agent and serves as a measure of success. An agent is deemed more intelligent if it consistently selects actions that yield outcomes better aligned with this function. The function can be simple, such as assigning +1 for a win and -1 for a loss in a game of Go, or complex, requiring the agent to evaluate and learn from past actions. It encapsulates all goals and incorporates trade-offs between conflicting objectives, as seen in a self-driving car balancing safety, speed, and passenger comfort.
Different contexts use various terms for this concept. In economics and decision theory, it is a utility function; in optimization, an objective function; in Machine learning, a loss function; in Reinforcement learning, a reward function; and in evolutionary systems, a fitness function. Goals can be explicitly programmed or induced through learning or evolution. For example, a reinforcement learning agent learns to maximize cumulative reward, while an evolutionary algorithm's behavior is guided by a fitness function that determines which agents reproduce, analogous to natural selection.
Agent-Based Definition of AI
Russell and Norvig's framework describes AI as the study of agents that receive percepts from an environment and perform actions. In this view, an agent is anything that perceives through sensors and acts through actuators. A rational agent selects the action expected to maximize its performance measure, given its percept sequence, prior knowledge, and available actions. This definition emphasizes that goal-directed behavior is central to intelligence, and it applies to both simple and complex systems.
Agent-oriented computing further characterizes agents by autonomy, responsiveness, and proactive behavior. The BDI architecture models practical reasoning by representing an agent's beliefs about the world, desires (objectives), and intentions (committed courses of action). This approach is widely used in multi-agent systems and robotics, where agents must operate in dynamic environments.
Types and Examples
Intelligent agents can be classified by their complexity and learning capabilities. Simple reflex agents act solely on current percepts, while model-based agents maintain an internal state. Goal-based agents consider future actions to achieve objectives, and utility-based agents maximize a utility function. Learning agents improve their performance through feedback, such as reinforcement learning or evolutionary computation.
Examples include a basic thermostat, which is a simple reflex agent, and AlphaZero, a chess program with a simple objective function of +1 for a win and -1 for a loss. More complex agents include self-driving cars, which balance multiple objectives, and software agents that carry out tasks on behalf of users, such as Waymo's autonomous vehicles or Tesla systems. In economics, agents are closely related to rational agents, and the paradigm is studied in cognitive science, ethics, and philosophy.
Learning and Adaptation
Agents often improve through learning algorithms. In Reinforcement learning, a reward function provides feedback, encouraging desired behaviors and discouraging undesirable ones. Programmers may use reward shaping to give incremental rewards for progress, as noted by Yann LeCun in 2018: "Most of the learning algorithms that people have come up with essentially consist of minimizing some objective function." Evolutionary computing can evolve agents that appear to act intentionally, guided by a fitness function.
Some systems, like nearest-neighbor classifiers, reason by analogy rather than being explicitly goal-driven, but they can still be framed as agents with implicit goals in their training data. Similarly, knowledge-representation systems can be viewed as agents with the goal of answering questions accurately, extending the concept of "action" to include providing answers. In generative adversarial networks (GANs) of the 2010s, a generator mimics human text composition, maximizing a function that represents fooling a discriminator, illustrating mimicry-driven goal functions.
Applications and Implications
AI agents are deployed across industries, from autonomous vehicles and robotics to virtual assistants and recommendation systems. Companies like OpenAI, Anthropic, and Google DeepMind develop agent-based AI systems, while Amazon Web Services and Microsoft Azure offer cloud infrastructure for deploying them. The paradigm also influences research in cognitive-science and socio-cognitive modeling, where agents simulate social behaviors.
The design of objective functions raises ethical considerations, as agents may optimize for unintended outcomes. For instance, a self-driving car's objective must carefully balance safety and speed. As agents become more autonomous, ensuring alignment with human values becomes critical. This has led to research in AI safety and interpretability, with institutions like MIT CSAIL and Stanford AI Lab exploring these challenges.
Future Directions
Advances in Large language models and Transformer (architecture) architectures have enabled more sophisticated agents that can reason and plan over extended periods. These agents, often called agentic AI, proactively pursue goals and make decisions in dynamic environments. Research continues on improving learning efficiency, robustness, and generalization, with contributions from academic labs like BAIR (Berkeley AI Research) and Carnegie Mellon University.
The evolution of AI agents will likely lead to more integrated systems that collaborate with humans and each other, raising questions about accountability and control. As of the early 2020s, agent-based AI remains a vibrant field, with ongoing debates about the nature of intelligence and the role of objective functions in shaping behavior.