Glossary

A software system that perceives its environment, reasons about a goal, takes actions using tools, observes the results, and adapts — repeating until the goal is reached or it gives up.

AI that has autonomy — it can plan and act toward a goal without a human directing every step.

The cycle every agent runs: Think (reason about the goal) → Act (use a tool) → Observe (see what happened and update memory) → repeat.

The instructions or question given to an AI.

A neural network trained on vast text data to predict what words come next — the "brain" that powers modern AI agents.

The basic unit an LLM processes — roughly a word or part of a word.

When an AI confidently states something false because it generates plausible-sounding text, not necessarily factual text. A mechanism — distinct from misinformation, which is the harm caused when someone acts on a wrong answer.

Prompting an AI to "think step by step" before answering, which improves accuracy on complex tasks.

Background instructions given to an AI before the user interacts with it — it shapes how the agent behaves.

Zero-shot: no examples given. Few-shot: examples included in the prompt to guide the response.

An agent's ability to invoke an external function — search, calculator, code runner, database query — and use the result in its response.

The amount of text an LLM can "see" at once — its short-term working memory.

Fetching relevant documents and giving them to the LLM so it answers from real information, not just training data. Like an open-book test for the AI.

A specialized database that stores information as mathematical representations (embeddings), enabling similarity-based retrieval.

Information stored outside the LLM and retrieved across multiple conversations.

A network of AI agents that collaborate to complete complex tasks.

The "manager" agent that breaks down a task and assigns subtasks to other agents.

An agent with a specific skill — coding, writing, research, verification.

The passing of information or control from one agent to another.

Rules or checks that prevent an agent from taking harmful or unauthorized actions.

A design pattern where a human reviews or approves agent actions before they execute — required before anything irreversible or affecting someone else.

A specific way an agent can go wrong. Identifying failure modes — and how you'd catch them — is part of building responsibly.

Systematic errors in AI output that reflect unfairness in the training data.

An attack where malicious text tricks an AI agent into ignoring its instructions.

The challenge of making AI systems pursue the goals their builders actually intend.

The field of research dedicated to ensuring AI systems behave reliably and beneficially.

A set of principles — fairness, transparency, accountability, privacy — for ethical AI development.