An AI agent receives an objective and determines the steps to achieve it, invoking tools, evaluating results, and adjusting its approach. It differs from a single-request system by operating across multiple turns without step-by-step direction.
Agents typically operate within a defined set of available tools and permissions.
Alternative Names:
Autonomous Agent, Software Agent
Why it Matters?
Agentic operation changes the supervision question from reviewing an output to reviewing a process, because the system took actions along the way that produced the result. For legal work the practical controls are limiting available tools to read operations, requiring approval before anything with external effect, and maintaining a complete action log. An agent that can file, send, or delete without confirmation is a different risk category from one that can only read and report.
Frequently Confused with
Related terms
Frequently asked questions
What distinguishes an agent from a chat interface?
What controls suit legal use?
