What is an Agent? Definition & Scope
An LLM-based 'agent' = autonomous system that perceives, decides, acts, observes.
In classical AI: an agent perceives via sensors and acts via actuators. Russell & Norvig: rational agent — one that maximizes expected utility.
LLM-based agent: uses an LLM as its "brain", interacts with the world via tools, autonomously solves multi-step tasks. Different from a chatbot: chatbot = single-turn; agent = multi-turn + tools + memory + planning.
OODA Loop: Observe → Orient → Decide → Act. LLM agents mimic this: read prompt → reason → pick tool → execute → interpret result → reason again.
3 critical traits: (1) Autonomy — proceeds without user-step-by-step guidance, (2) Tool use — connects to the external world, (3) Adaptation — revises plan based on observations.
Why this matters
Foundation stone for the shared vocabulary of all agent literature.
What you'll gain
You can tell whether a system is truly an 'agent', and defend why agent (or workflow) is the right choice.