I build

From LLM to agent in 20 terms

The climb through the first three floors, in the order in which you meet them while building. You start from what a model can do, install around it what it lacks, then put it in a loop. The end of the path is about what decides whether the system can go to production.

Engineers, architects, technical product people · 20 terms · 45 minutes · final quiz

The sequence

0 / 20 seen

Large language modelThe starting point: a function that, from a text, produces the most probable continuation. No state, no memory, no side effects.
TokenThe real unit of processing and of billing. Everything you design is measured in tokens before it is measured in euros.
Context windowThe hard constraint of floor 1. It bounds what the model can see in one call, answer included, and it is re-read in full at every turn.
HallucinationThe direct consequence of predicting what is plausible. You do not remove it: you set bounds around it, by what you give it to read and by what you check.
Fine-tuningThe only lever that really touches the model. To be kept for behaviour, never for knowledge that changes.
HarnessThe move up to floor 2: everything built around the model. It is the bulk of your code, and the bulk of your incidents.
System promptThe permanent instructions, sent again with every request. They take up context, they cost on every call, and they are not secret.
ContextWhat the harness assembles before every call. The quality of a system owes more to that assembly than to the choice of model.
RAGSelecting the right passages rather than sending everything. It is an answer to the window, to cost and to traceability, in that order.
Vector databaseThe means of retrieving by similarity of meaning. It serves RAG, it does not constitute it: the quality comes from the chunking and the ranking.
ToolWhat the system is allowed to do outside of text. Every tool added is a capability, and a surface of risk.
Tool callingThe exact mechanism: the model asks, the harness executes. The model never touches anything itself, and that boundary is your point of control.
Structured outputConstraining the shape of the answer to make it usable by code. Constraining the shape guarantees nothing about the substance.
GuardrailsWhat limits the damage, upstream and downstream. To be designed at the same time as the capability, never after the incident.
AgentThe move up to floor 3: the model chooses the next step. You lose predictability and gain adaptation.
Agentic loopThe mechanism: observe, decide, act, take note, start again. Since the context is never reset, the cost grows faster than the number of turns.
WorkflowThe alternative that gets dismissed too quickly. When the steps are known, a written sequence remains safer, cheaper and auditable.
AutonomyA dial you set risk by risk, never a switch. The useful question is: what can it do without validation?
EvalsWhat replaces intuition once the system stops being deterministic. Without a set of evals, an improvement cannot be told apart from a regression.
ObservabilityWhat makes it possible to understand, after the fact, why it did that. On a system that chooses its own steps, it is a design requirement.

Your progress stays in this browser. No account, no data sent.

What to remember

The summary

The bulk of your work is on floor 2

The model is a component you almost never modify. What you write, maintain and debug is the assembly of the context, the exposure of the tools, the constraint on outputs and the guardrails. A project that thinks of itself as a model project has the wrong subject.

The context is your main lever, and your main cost

It is re-read in full at every call: its size is paid for as many times as there are turns. That is why selecting beats stacking, why a long conversation becomes expensive, and why an agent that chains ten steps does not cost ten calls, but more.

Only go to an agent if a workflow will not do

An agent is justified when the steps are not known in advance. Everywhere else, a written sequence is safer, cheaper, replayable and auditable. The choice is made use case by use case, and it is perfectly legitimate for a workflow to contain an agentic step.

Without evals or traces, you are steering nothing

A system that is not deterministic cannot be validated by eye on three examples. You need a measurable set of cases to tell an improvement from a regression, and usable traces to reconstruct a sequence after the fact. Both pieces of work start before going to production, not after the first incident.

The final quiz

3 questions · click your answer

Level 2 · Distinguish

An internal assistant answers well on recent documents and badly on older ones. The corpus is complete and up to date. What do you look at first?

Level 2 · Distinguish

You expose a tool that sends emails to an agentic system. Where do you place the control?

Level 1 · Recognise

Why can the cost of an agent not be deduced from the cost of one call multiplied by the number of steps?

v2026-08 · pathCite this path