What is an AI agent?
An AI agent is software given authority to interpret inputs and choose tool calls or multi-step actions toward a goal.
How it works and where it is used
A language model may suggest plans, read data and act through defined APIs. Not every automated workflow is autonomous. Tool output can be wrong or malicious, and a model may invent results. Distinguish permission to read from permission to act.
Limits, safety and selection criteria
Give narrow permissions, require human review for critical actions and plan reversibility. Limit budget, time, retries and data exposure. Measure success by verified outcomes rather than fluent explanations. An AI agent can plan steps, use tools and adapt to intermediate results. A travel agent might find routes, compare times and draft an itinerary. Changes to calendars, bookings or payments need scoped permission and final confirmation. Text returned by tools is untrusted input; instructions embedded in a web page should not override the user's request. Keep a record of sources, tool actions and failures. Narrow permissions and human review matter most when the agent can change external systems.