(noun)
A software system that perceives its environment, makes decisions, and takes actions autonomously to achieve a defined goal. What separates an agent from a system that simply answers a prompt is that an agent operates over a sequence of steps — observing, reasoning, acting, then observing the result and adjusting — while holding onto its objective across the whole sequence.
Why the distinction matters. The word gets stretched to cover almost anything with AI in it, but the useful boundary is the perceive–decide–act loop. A system that produces one output and stops is not exercising agency; it’s responding. An agent earns the label by carrying a goal forward and making its own choices about what to do next. The same underlying model can be an agent in one setup and not in another — the difference is whether it controls a sequence of actions or just fills in a single blank.
The environment is central. An agent’s environment might be a file system, a web browser, an API, a codebase, or even other agents. What it can perceive and what actions are available to it are defined entirely by that environment. This is why the same model behaves very differently depending on the tools and context it’s given — the environment sets the boundaries of what agency is even possible.
Example. A coding agent receives a bug report. It reads through the codebase to locate the fault, edits the relevant files, runs the test suite, sees which tests still fail, and revises its fix — looping until everything passes. At no point does a human tell it “now run the tests” or “now try a different approach”; it perceives each result and decides the next move itself.
Common misconception. That an agent must be a single, neatly bounded program. In practice an “agent” is a role a system plays, not a fixed architecture — what matters is the autonomous, goal-directed loop, not how many files or models implement it.
Related terms: agentic AI, workflow, pipeline, multi-agent system
