Practice: What makes an AI an "agent"
Self-check
Section titled “Self-check”Seven short questions. Answer each in your head before opening the collapsible. Active retrieval is where the learning sticks.
1. In one sentence, what makes a system an agent rather than a chatbot?
Show answer
A loop with tools. A chatbot is one shot: text in, text out, no way to act in the world. An agent wraps the same model in a loop where it can call tools, read the results, and keep going until the goal is met.
2. Name the three repeating moves of the agent loop.
Show answer
Perceive (read the current state: the goal plus results so far), decide (choose the next move, often a tool call), act (run the action, observe the result, feed it back into the next perceive).
3. What are the four parts that make a system agentic?
Show answer
The model (the decision-maker), the system prompt (tells the model it may call tools and how), the tools (functions that reach the world), and the loop (outside code that runs the tools and feeds results back). The model is one of four parts, not the whole thing.
4. Fill in the blank. “The agency lives in the ______; the intelligence lives in the ______.”
Show answer
Scaffolding (system prompt + tools + loop); model. Swapping in a stronger model improves the decisions but does not change whether the system is an agent. Agency is a property of the surrounding scaffolding, not the model.
5. Why can a plain chatbot not answer “what is the weather in Seattle tomorrow?”
Show answer
It has no tools and no loop. It can only produce text from what it learned in training; it cannot reach outside that to check live data. The conversation is a sealed room. An agent breaks the room open by giving the model a weather tool and a loop to call it.
6. When is an agent the wrong choice?
Show answer
When the task is a single fixed step: one lookup, one translation, one summary. The loop costs latency, money, and reliability. For a one-shot task a plain model call is faster and more dependable. An agent earns its cost on open-ended, multi-step, or improvable tasks.
7. Someone says, “We switched to a much stronger model, so now our chatbot is an agent.” What is wrong with that?
Show answer
Model strength has nothing to do with whether a system is an agent. A stronger model makes better decisions, but it is still just predicting text. The system becomes an agent only when you wrap it in a loop, give it tools, and tell it (via the system prompt) that it may call them. Agency comes from the scaffolding, not from a better model.
Try it yourself: agent or single call?
Section titled “Try it yourself: agent or single call?”No tooling, no cost; this is judgment practice. For each task, decide whether it is better served by a single model call or by an agent (a loop with tools), and say why in one line. Then check.
A. "Summarize this paragraph I pasted."B. "Watch my inbox and draft replies to any scheduling requests."C. "Translate this sentence into French."D. "Research three competitors, pull their pricing, build a comparison table."Show answer
- A: single call. One-shot transformation of text already in front of the model. No tool, no loop needed.
- B: agent. Open-ended and ongoing; needs tools (read inbox, draft mail) and a loop that runs over time. Not a single fixed step.
- C: single call. One-shot transformation, like A.
- D: agent. Multi-step (several lookups), needs a tool to fetch each competitor’s data, and the steps cannot be fully scripted in advance.
The test: a single fixed step that transforms text already present is a call; anything that must reach the world, repeat, or adapt across steps wants an agent.
Try it yourself: trace the loop
Section titled “Try it yourself: trace the loop”Write out a perceive-decide-act trace for this task, the way the lesson traced the meeting-booking example. Assume the agent has the tools check_stock(item) and place_order(item, qty).
Task: “Reorder printer paper if we are running low.”
Show a sample trace
PERCEIVE: goal is reorder printer paper if low; no results yet.DECIDE: I need to know the current stock. Use check_stock.ACT: check_stock("printer paper") -> { units: 2 }PERCEIVE: only 2 units; that is low.DECIDE: Low enough to reorder. Place an order.ACT: place_order("printer paper", qty=10) -> { confirmed: true }PERCEIVE: order confirmed.DECIDE: Goal met. Report to the user.ACT: reply: "Stock was low (2 left); ordered 10 more."Yours does not need to match word for word. What matters: you perceived before deciding, each decision led to one action, and the result of each action fed the next perceive. If “2 units” had been “plenty,” the correct trace ends early with no order, which is itself a valid outcome of the decide step.
Flashcards
Section titled “Flashcards”Ten cards. Click any card to reveal the answer. Use the Print flashcards button to lay out the full set as one card per page for offline review.
Q. What is an agent, in one sentence?
A model wrapped in a loop: 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.
Q. What are the three moves of the agent loop?
Perceive (read the current state), decide (choose the next move), act (run it, observe the result, feed it back).
Q. What is the single difference between a chatbot and an agent?
The loop (and the tools it uses). A chatbot is one shot, text in and text out, sealed off from the world. An agent reaches outside through tools and iterates.
Q. What four parts make a system agentic?
The model (decides), the system prompt (grants tool use and sets the call format), the tools (reach the world), and the loop (runs tools and feeds results back).
Q. Where does 'agency' come from if the model is just predicting text?
From the scaffolding around the model: the system prompt, the tools, and the loop. The intelligence lives in the model; the agency lives in the scaffolding.
Q. If you swap the model in an agent for a stronger one, what changes?
Only the quality of the decisions. The system prompt, tools, and loop are untouched, so the agent keeps the same shape and powers; it just decides better.
Q. Why can a plain chatbot not report tomorrow's weather?
It has no tools and no loop, so it cannot reach live data. It can only produce text from training. An agent adds a weather tool and a loop to call it.
Q. When should you NOT use an agent?
For a single fixed step (one lookup, one translation, one summary). The loop’s cost in latency, money, and reliability is not repaid. Use a plain model call.
Q. When does an agent earn its cost?
When the task is open-ended (cannot be fully scripted), multi-step (several tool calls across turns), or improvable (it should use results to do better as it goes).
Q. What changed about agents when language models arrived?
The ‘decide’ step. Older agents decided with hand-coded rules or explicit search; a language model decides from what it learned in training, which is why agents now handle messy, open-ended tasks.