Agentic AI

A broad label for AI-enabled systems that select steps, use tools, observe results, and continue toward a goal across more than one action.

Agentic AI covers interactive assistants with tool loops, unattended agents, and model-directed stages inside larger workflows. An email system might retrieve a thread, draft a reply, check a calendar, and request a follow-up event. The operational significance comes from its tools, state, and authority, not from the label. A one-step model call can have a large external effect, while a long research loop can remain read-only.

Builder example

Prompts are only one control surface. Typed tools, scoped credentials, transaction limits, idempotency, external validation, event records, stop conditions, and recovery determine how generated decisions become effects. Choose those controls from the task's reach and failure modes.

Common confusion: Agentic does not mean fully autonomous, and autonomy is not one linear scale. A workflow can have broad latitude to read, narrow permission to write, and an independent authorization path for one operation.