Intelligent agent

In artificial intelligence, an intelligent agent is a system that observes its environment, makes decisions, and takes actions to achieve specific goals. It may improve its performance by learning or using stored knowledge. AI is often defined as the study and creation of these agents, focusing on their ability to act rationally toward goals.

A subset called agentic AI (or AI agents) can independently pursue complex goals over time, making decisions without constant human input.^[1]

Intelligent agents vary from simple systems, like a thermostat, to complex ones, like a self-driving car or a human being. They work based on an objective function that defines their goals, aiming to maximize its value through planned actions.^[2] For example, a reinforcement learning agent uses a reward function to guide its behavior.^[3]

These agents are studied in fields like economics, cognitive science, and ethics, and are related to software agents—programs that perform tasks for users. They are sometimes called rational agents, a term from economics.^[4]

Agents are classified by their complexity and decision-making abilities:

1. Simple Reflex Agents: Act based only on the current input, using "if-then" rules. Example: A thermostat turning on when it’s cold.^[5]

2. Model-Based Reflex Agents: Keep an internal model of the environment to handle partial information. They use past inputs to make decisions.

3. Goal-Based Agents: Choose actions to reach specific goals, often using search or planning. Example: A Roomba vacuum navigating a room.^[6]

4. Utility-Based Agents: Evaluate actions based on a utility function, measuring how desirable outcomes are. They aim for the best possible result.

5. Learning Agents: Improve over time by learning from feedback. They have a learning part that adjusts actions based on performance.

Agentic AI refers to advanced agents that operate autonomously, often using large language models (LLMs). They can handle tasks like booking travel or coding without constant human guidance.^[7] Examples include Devin AI, AutoGPT, and OpenAI Operator.^[8]

Agentic AI can: - Boost productivity by automating repetitive tasks.^[9] - Support innovation by handling complex problems. - Improve web accessibility for people with disabilities.^[10]

Challenges include: - Privacy risks and cybercrime.^[11] - Ethical issues and AI safety concerns. - Potential job displacement or algorithmic bias.^[12]

• Self-Driving Cars: Use agents to navigate roads, tested in simulations like Waymo’s Carcraft.^[13]
• Business: Salesforce’s Agentforce builds agents for tasks like customer service.^[14]
• Security: The Transportation Security Administration uses agents for biometric authentication.^[15]

• Reinforcement learning
• Software agent