In depth
The four parts of an agent are the model, the role prompt, the tools, and the loop. The model is the reasoning engine, usually a current Claude, GPT, or Gemini variant. The role prompt is the agent's job description: who it is, what it owns, what it never does. The tools are the integrations it can call, which might be a Gmail account, a CRM API, a code sandbox, a Slack workspace, or a vector database. The loop is what makes it an agent rather than a one-shot completion. It can take an action, read the result, decide on the next action, and keep going until it is done or it hits a stopping condition.
What an agent is not: a single chat response with extra steps, a Zapier workflow with a language model bolted onto one node, or a fully autonomous employee that needs zero oversight. The most useful agents today are narrow. One role. A handful of tools. A clear definition of done. A human or another agent checking the output where the stakes are high.
The common confusion is between an agent and an assistant. An assistant waits for you to ask a question and then answers it. An agent has standing work it owns, like drafting weekly LinkedIn posts, triaging inbound leads, or reconciling invoices against the bank ledger, and runs that work on a schedule or in response to an event. The same underlying model can power both. The difference is the operating context wrapped around it.