AI agent

An AI agent is a system in which a language model is given tools — searching, reading a file, calling an API, writing to a database — and allowed to decide which to use, in what order, until a task is done. The distinguishing feature is not the model but the loop: it acts, sees the result, and chooses again.

Agents are genuinely useful for multi-step work with an uncertain path, such as investigating a support ticket across several systems. They are risky in proportion to what their tools can do, because an incorrect step is not just a wrong sentence but a real action taken on your infrastructure.

Sound practice is to scope tools tightly, make destructive actions require confirmation, log every step the agent takes, and cap how many iterations it may run before it has to hand back to a human. An agent with unbounded tools and unbounded patience is an outage waiting for its moment.