AI agent
Also called: agent, agentic AI, autonomous agent
An agent is an AI you hand a goal to that then decides the steps and takes the actions itself — sends the email, updates the record, moves the money — without stopping to ask permission at each one.
An agent is an AI you hand a goal to that then decides the steps and takes the actions itself — sends the email, updates the record, moves the money — without stopping to ask permission at each one. That last clause is the whole definition, and it’s the exact line a sales pitch will blur: a system that produces a draft and waits for you is a tool wearing an agent’s name tag.
Why it matters: Agents are the most oversold idea in AI, and it isn’t close. What ships under the label is usually one of two things: not really an agent (it hands you text and waits), or a real agent that isn’t yet trustworthy enough to hand anything that matters. The quiet reason is math — a system that must be right at every step of a long chain compounds its errors, and at 95% reliability per step, a ten-step chain succeeds fifty-nine percent of the time. Knowing the strict definition is how you avoid paying agent prices, and carrying agent risk, for what is functionally a chatbot.