What is an AI agent?

Point a language model at a goal and let it work out what to do next, one step at a time, and you have crossed the line from automation into agency. That single property, who authors the sequence of steps, is the cleanest way to tell an agent apart from a program that merely has AI wired into it somewhere.

What makes something an agent

The defining trait is where the control flow comes from. In an ordinary program, a developer writes the branch logic ahead of time: if this, call that; then this; loop back here. The software follows a path someone already drew. An AI agent inverts that. It is handed a goal and works out its own path as it goes, deciding each next action from the situation in front of it rather than from a script.

A system where a person wrote the branch logic and the model only fills in a blank is not an agent, however much AI sits inside it. A system where the model reads the current state and picks the next move is.

The loop that does the work

Underneath, an agent runs a loop: observe the current situation, reason about what to do, take an action (often by calling a tool or another system), read the result, and go again until the goal is met. The shape of that cycle predates modern AI. What changed recently is the reasoning step. In ReAct, Yao and colleagues showed a language model generating its reasoning and its next action interleaved in the same stream, so the model itself issues the action to take next instead of a fixed controller choosing it (Yao et al., 2022).

That is the mechanical basis of runtime control-flow authorship. The “choose an action” step, historically written by an engineer, is now performed by the model in-line with its own reasoning.

Why the idea is older than large language models

The word agent is not fresh marketing. It is the standard frame for the whole field of AI. The canonical textbook, Russell and Norvig’s Artificial Intelligence: A Modern Approach, describes an agent as anything that perceives its environment through sensors and acts upon it through actuators, and casts AI itself as the study and design of rational agents that pursue a performance measure. A thermostat fits that definition. So does a chess program.

What large language models added was a general-purpose way to perform the deciding step across open-ended tasks stated in plain language, instead of tasks hand-coded one at a time. The definition did not move. The range of goals a single system can chase under it did.

How this differs from a chatbot

Because an agent acts and does not only answer, it sits apart from a chatbot, which replies inside a conversation and then waits for you. That contrast has an axis of its own worth reading separately; see AI agent vs chatbot. For the definition itself, the deciding line is not fluency or how human the replies sound. It is whether the system chooses its own next step.

The part most definitions skip

Everything else people attach to the word, tools, memory, autonomy, multi-step planning, follows from that one property. Give a system the right to choose its next step and it needs tools to act with and memory to track where it stands. Take that right away and what remains is a program with a model bolted into a single slot.

The practical consequence is concrete: two runs of the same agent on the same goal can take different paths, because the path is written at runtime and never fixed beforehand. When you want to know whether something is really an agent, look at the last action it took and ask who chose it. If a developer wrote that choice into the code, you have automation. If the model made the choice from the goal, you have an agent.

Sources

• Russell, S. and Norvig, P. Artificial Intelligence: A Modern Approach. 1st ed. 1995; 4th ed. 2020.
• Yao, S. et al. ReAct: Synergizing Reasoning and Acting in Language Models. arXiv:2210.03629, 6 Oct 2022.