Intelligent Agents and Their Capabilities

The word agent appears all over engineering. You may meet software agents, intelligent agents, user agents, conversational agents, and reinforcement learning agents. Each name points to a different kind of system, so it is easy to wonder whether these are unrelated ideas that just share a label. Before comparing them, it helps to settle on the base idea that the term rests on, because that base idea is what lets the word be used at all.

So the question for this article is simple: what makes something an agent in the first place? The source gives a short, direct answer. An agent is anything that can perceive its environment and act upon that environment. Every system in the list above can be analyzed with this definition, which is why it is the reliable starting point.

Notice that the definition names two separate abilities and joins them with and. Perceiving means taking in information from the environment. Acting means doing something that affects that environment. The definition is not satisfied by one ability alone. An agent is characterized by its ability to perceive and to act, so both must be present together.

The small word upon matters too. It ties the two abilities into one loop: the agent acts on the same environment it perceives. The agent does not sense one world and change a different one. Its actions land back in the surroundings that its perception reads from.

Read the definition as a path rather than a picture of parts. Information leaves the environment, reaches the agent as perception, and the agent responds with action that reaches back into the environment. Because the action changes the environment, the next perception the agent receives is shaped by what it just did. The path is therefore a loop, not a straight line.

The loop explains why the two abilities cannot be separated. Perception without action has no way to change what comes next. Action without perception has no information about the surroundings to guide it. The definition asks for both because the loop needs both halves to close.

Consider a simple heater controller in a room. You do not need any special vocabulary to describe it. The steps below are an original illustration written for this article, not taken from the source. Trace them in order and check each step against the two requirements from the definition.

Now change one thing. Suppose the controller only displays the temperature and never switches anything. The perception is still there, but the action is gone. Before you read the answer, decide what the definition says about this version.

The source sets the two abilities side by side because they answer different questions. Perception answers what the system can learn about its surroundings. Action answers what the system can change in those surroundings. Keeping the questions apart makes it easier to see which half a system is missing.

Only the heater controller in the diagram meets the definition. The thermometer display perceives but never acts. The blind switch acts but never perceives its surroundings, so it has no information to act upon. Both of these systems are missing one half, and by the definition neither counts as an agent.

A second way to see the two abilities is as flows in opposite directions. Perception carries information inward, from the environment to the agent. Action carries the agent's effects outward, back into the environment. The diagram below shows this as a path that starts and ends in the same place.

When you analyze a system, you can ask two separate questions. Which of its behaviours are inputs from the environment? Which of its behaviours change the environment? Each behaviour falls into one of these groups, and an agent needs at least one behaviour in each.