What is an Agentic AI?

The defining trait of agentic AI is a loop: the system observes its current state, reasons about what to do next, takes an action (often a tool call), observes the result, and repeats until the goal is met or it decides to stop. This is a fundamentally different shape from a classic chatbot turn, which ends after one response regardless of whether the underlying task is actually finished. Agentic systems introduce new failure modes that pure text generation doesn't have: an agent can get stuck in a loop, take an irreversible action based on a wrong assumption, or drift off-task across many steps. This is why agentic platforms invest heavily in sandboxing, tool permissioning, tracing, and guardrails, controls that matter far less for a single-turn assistant.