AI Agent
An AI agent is a system built around a model that decides its own next action, step by step, until a task is finished. At each step, the model looks at where things stand and picks what to do next — call a tool (a specific outside action it can trigger, like running a search or reading a file), look something up, ask for more information, or stop because the task is done. Nobody wrote that exact sequence of steps in advance. The model works it out each time the program actually runs.
That's the one thing that makes something an agent instead of a plain script: who decides the next step. In a normal program, a person wrote every step ahead of time. In an agent, the model decides while the program is running, based on what just happened.
Agent or workflow?
A single call to a model takes one input and produces one output — it can't look something up partway through and change its answer based on what it finds. A fixed workflow can call tools and look things up too, but a person has to write out every possible path in code ahead of time. That works well when the steps are known and don't change.
An agent is for the case in between: you know the goal, but not the exact steps to reach it, or the right steps depend on what happens along the way. Debugging an unfamiliar error is a good example — read the error, check the logs, form a guess, look at the code, try a fix, run the tests. What happens at each step depends entirely on what the last step turned up. A person can't write that branching logic out in code for every possible error in advance, so the model works it out live instead, one decision at a time.
If you can already write out the exact sequence of steps — even a long one — a fixed workflow beats an agent every time: it's more predictable, cheaper to run, and easier to debug, because nothing is left for the model to get wrong. That's also the field's own working advice, not just this site's: Anthropic's engineering team and OpenAI's docs both say to start with the simplest approach — often a fixed workflow, or even one plain model call — and only reach for an agent's open-ended decision-making once you genuinely need it. Every decision an agent makes is a place it can go wrong; a workflow with ten fixed steps has zero such places, while an agent making ten decisions has ten.
Related but different terms
"Agentic" is a spectrum, not a yes-or-no label — a system can use a model to make a few decisions inside an otherwise fixed process without being a full agent. "Agentic workflow" usually means exactly that: a mostly-fixed process with some model-made decisions mixed in, short of full autonomy.
A subagent is a full agent in its own right — the difference is only who calls it: another agent, instead of a person.