In plain terms
A chatbot answers; an agent acts. Ask a chatbot to “find three suppliers and compare their prices” and it tells you how. An agent searches, opens the sites, builds the table and brings it back. The loop makes the difference: act, check, decide what comes next.
Why it matters
Agents move AI from answering questions to finishing work, which is where most of the productivity gain sits. The risk profile changes too: a wrong answer is a bad paragraph, a wrong action is a sent email or a deleted record. Scope, permissions and oversight become design decisions.
Example
A coding agent receives a bug report. It reads the code, runs the tests, edits a file, reruns the tests and opens a pull request once they pass: a dozen tool calls with no human in between.
Most often confused with
Agent vs. Workflow
In a workflow a person fixes the steps and their order in advance; the model does its assigned job at each step. In an agent the model decides the next step. Workflows are predictable and easy to audit; agents are flexible and harder to predict.
Under the hood
A common working definition: a model using tools in a loop. Components: the model, a tool set (APIs, code execution, search, often exposed through MCP), instructions, and a harness that manages context, memory and stopping conditions. “Agentic” is a spectrum, from fixed workflows with one model call per step to open-ended loops where the model picks every step. Errors compound as chains get longer and reliability drops, so evals, checkpoints and narrow permissions often matter more than the choice of model.