From model to agent
A chat model answers; an agent acts. The four parts that turn next-token prediction into a system that can change the world.
A chat model is a function: text in, text out. Ask it for the weather and it can only describe what a weather API would say. An agent is the same model wrapped in a loop that lets it call the weather API, read the result, and decide what to do next. Nothing about the model changed — what changed is that its output can now have consequences.
Start here
An agent is a model plus three things it doesn't have on its own: a set of tools it can call, a loop that runs until the goal is met, and state that persists across steps. The model still predicts the next token; the loop is what makes that token an action.
The LLM domain's Agents and tools lesson is the ten-minute overview of this idea. This domain is the deep version — we will build up the loop, tools, memory, planning, and the guardrails that make acting autonomously safe.
The loop in one picture
Agent loop — user: "What's a 15% tip on $84.50?"
step 0/7ReAct pattern: reason → act → observe → repeat. Illustrative trace; real agents choose tools the same way — by reading tool descriptions in the prompt and emitting structured calls.
A single turn — the model reasons, emits a tool call, the runtime runs it, the result comes back — is called a step. The agent repeats steps until it decides it is done. That repetition is the whole difference between a model that knows things and a system that does things.
The four parts
- The model — the reasoner. Usually an LLM, but it can be smaller, faster, or specialised (a decision model for routing, say). Its job is to choose the next action given the state.
- The tools — the hands. Anything with a typed interface: a calculator, a database query, a file write, an HTTP call, another model. A tool the model cannot call does not exist to it.
- The loop — the control flow. Reason, act, observe, repeat. It ends on success, on a step or cost budget, or when the model gives up.
- The state — the memory. At minimum the growing transcript of what has happened; often a scratchpad, a plan, or a store that outlives the context window.
The word is overloaded
"Agent" is used loosely, and the honest thing is to say so. The useful distinction is a spectrum of autonomy:
- a prompt chain runs a fixed sequence of calls — no model decisions about what comes next;
- a router lets the model pick which fixed path to take;
- an agent lets the model decide, step by step, how many steps to take and what each one is.
Careful
Most things marketed as agents are workflows: fixed graphs with a model inside a box or two. That is not a criticism — a workflow is usually the right answer (the last lesson in this chapter is about exactly this). But calling a pipeline an "agent" sets wrong expectations about cost, reliability, and what it can handle.
Check yourself
Eduspheria wiki · Agentic AI, Foundations
0 / 4 answered
Next: the loop itself — reason, act, observe, and when to stop.