An AI agent is a system that takes a goal and executes: it decomposes steps, calls tools (search, code, browser), reads results and corrects course. The key difference from a chatbot: a chatbot delivers an answer, an agent delivers a finished job.
The standard anatomy: an LLM as brain + a planning loop (typically ReAct: think, act, observe) + a toolbox + memory. ZCode, Claude Code, Devin and Computer Use all live in this family.
Design note: agents grow reliable by narrowing, not just by getting smarter. Tight scope + right tools + self-testing = dependable. Open scope + many tools = wandering.
