Summary: What makes an AI an "agent"

An agent is a model wrapped in a loop with tools. It looks at a goal, decides whether a tool would help, calls the tool, reads the result, and repeats until the goal is met or it gives up. That loop is the entire difference between an agent and the chatbot you already use. Everything else in this track (tool use, memory, planning, multiple agents, trust, security) is detail layered on top of that one idea. This summary is the scan-in-five-minutes version of the full lesson.

Core ideas

• A chatbot is one shot: text in, text out, once, with no way to act in the world. Ask it tomorrow’s weather and it correctly says it cannot check live data. It is a sealed room; it can only produce text from what it learned in training.
• An agent breaks that room open by wrapping the same model in a loop with three repeating moves: perceive (read the goal plus results so far), decide (choose the next move, often a tool call), act (run it, observe the result, feed it back). The loop repeats until done.
• The agent did not get smarter than the chatbot. Given a weather tool and a loop, it knew it could go get the weather; the chatbot had no such reach. Same model, different scaffolding.
• Four parts make a system agentic: the model (the decision-maker, predicting the next move), the system prompt (tells the model it may call tools and how to format a call), the tools (functions that reach the world: read a file, query a database, send mail), and the loop (outside code that runs tool calls and feeds results back).
• The agency lives in the scaffolding; the intelligence lives in the model. Swap in a stronger model and the agent decides better but keeps the same shape. There is no special “agent model”; the same model is agentic when it runs inside a loop with tools and is not when it does not.
• The power is in the repetition. On a multi-step task (book a meeting: check one calendar, check another, resolve a conflict, send the invite), each trip around the loop starts from a richer picture than the last, and the agent keeps going until the goal is actually done.
• The idea predates language models. Rule-based programs (ELIZA, 1960s), symbolic planners (STRIPS, 1970s), and reactive robotics (Brooks, 1980s) all perceived, decided, and acted. What changed with language models is the decide step: from hand-coded rules and explicit search to decisions generated from training, which is why agents now handle messy, open-ended work.
• Agents are not free. The loop costs latency, money, and reliability. Reach for an agent when a task is open-ended, multi-step, or improvable; for a single fixed lookup or one-shot generation, a plain model call is faster and more dependable.

What changes for you

Before this lesson, “AI agent” was a marketing phrase you could not pin down. Now you have a definition that holds up, and a test you can apply to any agent product you meet: find the four parts. What is the model, what is the system prompt telling it, what tools can it call, and what loop runs them? Once you can take an agent apart that way, the whole landscape stops being mysterious, and the rest of this track is just filling in how each part is built well.