What Your 'AI Agent' Is Actually Doing: 8 Terms Explained Simply

I keep seeing 'agentic' everywhere. As something that actually runs these workflows in production, let me explain the jargon with what it really means under the hood. Agent — a loop, not a brain. The model reads the state, picks an action, runs a tool, reads the result, repeats. That loop plus a goal is the whole 'agent.' No magic. Tool / function call — the agent's only hands. It can't click or type on its own; it calls an API. If there's no tool for it, the agent can't do it. Context window — short-term memory. Everything the agent 'knows' in the moment has to fit here. Overflow it and it forgets the start of the task. ReAct loop — think, then act, then look, then repeat. Most 'reasoning' you see is exactly this pattern, not deep thought. MCP — a standard so tools plug in like USB. One agent, many tools, no custom glue. Retry / backoff — where the time actually goes. A third of real runs is re-running because the page changed or the site blocked you. Idempotency — doing the same thing twice shouldn't break anything. This is the difference between a reliable agent and one that double-posts. Human-in-the-loop — the honest version of 'autonomous.' Real agents stop and ask for a code, a card, or a decision.