Glossary

Kept short. A definition that needs a paragraph has become a guide, and is linked as one.

Token

The unit everything is counted in. A chunk of text, roughly a common word — but code, punctuation and non-English text fragment into more tokens per visible character. More

Context window

Everything the model can see at once, measured in tokens. Re-sent in full on every turn, which is why long conversations get both expensive and worse. More

Effort

A control over how much thinking a model does on a request. Usually a cheaper lever than switching model, and the one people reach for second.

Agentic loop

A model repeatedly using tools and reading the results until a task is done. Tool results accumulate in the context and are re-sent every turn, which is where the cost of agentic work actually comes from.

Prompt injection

Instructions hidden in content the model reads as part of a legitimate task. Contained rather than solved: nothing distinguishes your instructions from the document's. More

Confabulation

A fluent, confident, wrong answer. Produced by the same process as the correct ones, which is why it looks identical and why confidence carries no signal. More

MCP

Model Context Protocol. A standard way to give a model access to external tools and data. Installing one grants real access, so it is a supply-chain decision.

Skill

A folder of instructions an agent loads only when a task matches. Its description is a routing rule rather than documentation, which is why most skills that never fire have a description problem. More

Deployer

Under the EU AI Act, someone who uses an AI system rather than placing it on the market. Most organisations are deployers, and deployer duties are lighter than provider duties. More

Session window

The shorter of the usage limits, resetting every five hours. It starts with your first request after the last reset, not at midnight. More