Module 016 min

What an agent actually is

A chatbot answers a question. An agent pursues an outcome. The difference is not intelligence — it's a loop and a set of hands.

After this module
  • You can name the four parts every agent has, and say which one is missing when something "isn't really an agent".
  • You can place any product you've heard of on a six-level autonomy ladder, from tool to self-improving.

Ask a chatbot "what's the cheapest flight to Chicago on Thursday?" and you get a confident paragraph, some prices it may have invented, and a link to go check yourself. Ask an agent the same thing and it opens a browser, searches three sites, compares the results, notices the 6am is cheaper but leaves you 12 minutes to connect, books the 8am, and puts the confirmation in your calendar — then tells you what it did.

Nothing magical happened in the second case. The same kind of model was used. What changed is that the model was wrapped in a loop, given tools, and allowed to keep state across steps. That wrapper is the whole discipline.

1.1The four parts

Strip away the branding and every agent — a coding assistant, a support bot that issues refunds, a research assistant that reads forty papers — is the same four-part machine.

ANATOMY OF AN AGENT Goal done = a checklist The model decides the next step no hands · no memory Tools search · code · files · APIs each one has a schema Memory / state what happened so far the running record The world — files, web, databases, people given call observe acts on repeat until done
The model is the smallest part of the picture. Most engineering effort — and most failures — live in the other four boxes: the goal, the tools, the state, and the loop that holds them together.
Model
The language model. It reads the current situation and predicts the next useful step. It cannot see, click, save or remember anything by itself.
Goal
A description of the finished state, specific enough to check. "Handle my inbox" is a wish. "Draft replies to the 6 unread customer emails and leave them unsent" is a goal.
Tools
The hands. A search API, a code sandbox, a calendar, a database, a browser. Each tool is a function with a name, a description of when to use it, and a strict input format.
Memory / state
The running record: what was tried, what came back, what is still outstanding. Without it, every step starts from zero and the agent repeats itself.
The trap

"Agent" is now a marketing word. If a product cannot take an action in the world and cannot recover from a failed action, it is a chat interface with a nice prompt. That is fine — it's just not an agent, and it will not do your taxes.

1.2The autonomy ladder

Autonomy is not one switch. It's a dial with recognisable stops, adapted from the same six-level scale used for self-driving cars. Drag the dial and notice what the human's job becomes at each level — that job is the real product decision.

Try it · autonomy dialL0 → L6

L0 · Tool

One function call. No memory, no decisions. A spell-checker. A summarize button.

HUMAN ROLE: operator · "I press the button."

Practical rule: most reliable production systems are L2–L3 with strong guardrails. L4 exists; L5 is mostly a pitch deck.

1.3Where the line sits

The same model, four different amounts of agency.
You have…It canIt cannotTypical build time
PromptTransform text you paste inLook anything up, remember anything, actMinutes
ChainRun a fixed pipelineChoose a different path when something is oddAn afternoon
RAG systemAnswer from your documents, with citationsDecide to do something about the answerA few days
AgentChoose steps, call tools, check results, retry, reportBe trusted blindly — it needs permissions and reviewA week to a quarter

Notice the build-time column. Every step up the ladder multiplies the number of ways the system can surprise you. That is why the honest advice, in every module after this one, is the same: build the simplest thing that could possibly work, then add agency only where it earns its keep.

Do this now · 3 minutes

Write down one task you did this week that felt repetitive. Then answer three questions: (1) What would "done" look like, as a checklist? (2) What tools would something need to touch — files, web, email, a database? (3) What is the worst thing that happens if it acts wrongly once?

Keep that note. It is the raw material for the challenges at the end, and it is a better starting point than any tutorial you'll find.

Check yourselfQuestion 1 of 1

A product lets you upload a PDF and answers questions about it in polished prose. Which part is missing before you could call it an agent?

60-second recap

  • An agent = model + goal + tools + state, held together by a loop. Remove any one and you have a chatbot, a script, or a demo.
  • Autonomy is a dial, not a badge. L2–L3 is where dependable systems live; each level up buys capability and costs predictability.
  • The wrapper is the engineering. The model is interchangeable; the goal definition, tool design, state handling and permission model are yours.