In plain language
Instead of following a fixed script, it plans, acts through connected tools, checks results, and adjusts.
Example
An agent reads a support email, looks up the order, checks the refund policy, and drafts a reply for approval.
Related terms
Large Language Model (LLM) · Tool Use (Function Calling) · Model Context Protocol (MCP) · Human-in-the-Loop
Common misconception
An agent is not always better than a fixed workflow; predictable, high-volume tasks are usually cheaper and safer as standard automations.
Implementation guidance
Give agents narrow goals and minimal tool permissions, require approval for irreversible actions, log every tool call, and evaluate results on real cases.