An AI agent is a system that uses a language model to decide which actions to take, in what order, to accomplish a goal, rather than following a fixed script or responding to a single prompt with a single answer. That decision-making loop (decide, act, observe the result, decide again) is what separates an agent from a chatbot or a simple AI feature.
What actually makes something an agent, not just an AI feature
It can take actions, not just generate text
An agent typically has access to tools: it can query a database, call an API, search the web, write a file, or trigger a workflow. A chatbot that only generates conversational responses, with no ability to do anything outside the conversation, isn't an agent by this definition, however capable the underlying model is.
It decides its own next step
Given a goal, an agent determines what needs to happen to accomplish it (which tools to call, in what order) and adjusts based on what each step returns. That's different from a fixed pipeline where a human predetermined every step; the agent makes the sequencing decision itself, within whatever boundaries it's been given.
It can react to what it learns mid-task
If a tool call returns an unexpected result, a genuine agent can change its next action based on that, rather than blindly continuing a predetermined script. This is what lets an agent handle real-world variability that a fixed automation script would break on.
What an AI agent is not
- Not just a chatbot. A conversational interface that only answers questions, with no tool access and no multi-step planning, is a chatbot, not an agent, even if it's built on the same underlying model.
- Not just automation. A fixed script that always performs the same sequence of steps regardless of what it encounters is automation, not an agent, even if an AI model is involved somewhere in that pipeline. We cover this distinction directly in AI agent vs. traditional automation.
- Not necessarily fully autonomous. Many well-built agents include a human-in-the-loop step for consequential actions. Requiring human confirmation at certain points doesn't disqualify a system from being an agent; it's a design choice about how much autonomy the agent has.
What agents are actually used for in production
Common real applications include customer support agents that look up account information and take defined actions, research agents that gather and synthesize information across multiple sources, and coding agents that write, test, and iterate on code with minimal step-by-step human direction. What they share is the same underlying pattern: given a goal, figure out and execute the steps needed to get there, adjusting along the way.
How we approach this
We build systems with the level of agentic autonomy the task actually calls for, from a simple AI feature to a fully agentic workflow with tool access and multi-step planning. We don't default to "agent" because the term is popular when a simpler system would do the job.