Agentic AI

AI that takes multi-step autonomous actions toward a goal, where the action is the output, actions chain, and speed can remove the human reaction window, requiring oversight controls enforced outside the agent's reach; the trustworthy controls are hard boundaries enforced below the AI, a plan-then-act gate (a human approves the whole plan before any step executes), and a kill switch the agent cannot rewrite.

An AI system that takes multi-step actions on a user's behalf, such as calling other software or executing transactions, with limited human review at each individual step. The newest and fastest-moving of the three AI categories, and the one most likely to fall into definitional gaps in older statutes.

AI systems that take actions in the world through tools, not just produce outputs for a human to act on. The defining property is agency: the system can choose and execute actions, so its mistakes are actions that already happened rather than recommendations you can decline.

AI that does tasks rather than only answering questions. It takes a goal, breaks it into steps, and carries them out, checking in rather than waiting to be told every move. Its arrival shifts the human skill from asking to directing and supervising.

AI tools that take a sequence of actions autonomously across multiple steps rather than answering a single prompt; they automate parts of the lower literacy layers but raise the stakes of a gap, because an early error propagates silently through later steps. The literacy skill they demand is checkpoint thinking: knowing where a human must verify and being able to audit what the agent did.