Glossary

Terms are listed alphabetically. Where a term was introduced in a particular chapter, that is where the fuller treatment lives.

Agent: A system that uses an AI model together with components such as memory, tools and an execution layer to work through a sequence of steps towards a goal, with some degree of independence and inside boundaries its designer set. The model is the reasoning engine inside it; most of the rest is ordinary software.
API (application programming interface): A structured way for one piece of software to ask another to do something. What turns an agent from a system that talks into one that acts.
Attention: The mechanism in a transformer that weighs how much each token matters for interpreting the others, which is how a model follows meaning across a passage.
Bias: Unfair or unbalanced behaviour a system may inherit from its training data, potentially producing skewed or discriminatory output.
Chain of thought: Prompting a model to set out its working before giving a final answer. The steps shown are a generated explanation you can audit, not a transcript of the model's internal computation.
Chunking: Splitting documents into passages before embedding them, so retrieval can return the relevant part rather than the whole file.
Context window: How much text a model can take into account at one time. A hard budget, and one of the main reasons multi-agent systems exist.
Control: A rule that actually constrains what a system may do, as opposed to a policy that describes what it should do. To count, it must be explicit, testable and owned by somebody. Controls are preventive, detective or responsive.
Coordinated team: An arrangement in which a lead agent divides one larger task among several specialists working at the same time, then combines what they return. The technical name for the lead is an orchestrator.
Embedding: A representation of text as a list of numbers encoding its meaning, so that similar meanings sit close together. What makes semantic search possible.
Fine-tuning: Training a pretrained model further on a smaller, domain specific dataset to specialise its behaviour.
Foundation model: A large model trained on broad data, able to perform many tasks without being rebuilt for each one.
Generative AI: Artificial intelligence that produces new content, text, images, audio or code, rather than only analysing or labelling what exists.
Guardrails: Explicit rules, filters, permission limits and review points defining what a system may and may not do. Instructions shape behaviour; guardrails constrain it — and a permission boundary is more reliable than an instruction.
Hallucination: Output that reads as confident and well-formed but is factually wrong or entirely invented.
Handoff: Any point where work passes from one agent to another. It may carry structured data, a summary, a reference, the complete result or task state. Every handoff is a place where information can be lost or misunderstood.
Hard blocker: A condition under which no readiness score is allowed to authorise anything: a prohibited use, an open safety finding, a required control that failed, or evidence too thin or too old to rely on.
Human in the loop: A design in which a person reviews or approves consequential actions before they take effect.
Human-augmented: A design in which the AI does the work and a person makes the consequential decision. Not technically a multi-agent system, and one of the most important arrangements in practice: the person is a designed part of the workflow, not an emergency fallback.
Large language model: A deep learning model trained on very large amounts of text, able to follow context, generate language, answer questions and summarise.
Least privilege: Granting an agent the narrowest permissions that still let it do its job. A constraint is more reliable than an instruction.
Long-term memory: Persistent storage letting an agent retain preferences and history across sessions. Built from external components, not supplied by the model.
Multi-agent system: A system in which several agents, each with a clear responsibility, work towards the same larger goal, coordinated by a router, an orchestrator or a fixed sequence. Dividing the work may buy parallel execution, specialisation, separate context windows, separate permission boundaries, an independent second judgement, or simply room for more information than one agent can hold. Size is only one of those reasons.
Multimodal: Able to work with more than one kind of data, such as text together with images, audio or video.
Neural network: A computational system of interconnected nodes that processes data and learns patterns. The comparison with the brain is a teaching analogy, not a description.
Orchestrator: A coordinating agent that plans a task, delegates parts of it, synthesises what returns and decides whether to continue. Distinct from a router, which only dispatches.
Parameters: The internal numerical values a model adjusts during training. A rough proxy for capacity, and frequently not published.
Pipeline: Agents arranged one after another, each transforming the output of the last. Easy to follow, and a mistake made early travels all the way down.
Pretraining: The initial phase in which a model learns general knowledge and language patterns from a large, varied dataset, before any specialisation.
Production: The environment where a system runs against real users, real data and real consequences, as distinct from development or testing.
Prompt: The text given to a model that tells it what is wanted: a question, an instruction or a description of the task.
Prompt engineering: Designing and structuring instructions deliberately to get more accurate and more useful output. Your prompt is the part you control; a working system usually assembles more around it, including system instructions, conversation history, retrieved documents, tool results and memory.
Prompt injection: A security risk in which text, often arriving from a retrieved document or web page, is read by the model as an instruction and overrides its intended behaviour.
RAG (retrieval augmented generation): Retrieving relevant external material and giving it to the model before it answers. Improves grounding and freshness; does not guarantee correctness.
Readiness: Whether an agent can be authorised to act in a particular place, at a particular level of independence, on the basis of evidence. Distinct from capability, which a demonstration establishes. The evidence required should rise with the authority granted.
Release gate: The point at which somebody decides whether an agent may be deployed. Three outcomes: approved, needs review, or blocked. Missing evidence is not a pass.
Router: A component that classifies an incoming request and sends it to the appropriate handler, model or specialist agent, without doing the work itself. Requests may be routed by subject or by difficulty. A router should always have somewhere to send the cases it cannot classify, which in practice means a person.
Router-based: An arrangement in which one component classifies each incoming request and sends it to the specialist best suited to handle it. The router does not do the work, and there is no step that combines results.
Short-term memory: Temporary state holding information within a single session or conversation.
Subagent: An agent spawned by another to handle a bounded task, with its own context window. Used for parallelism, for tool restriction, or to keep verbose work out of the main context.
Temperature: A setting controlling randomness in word selection. Lower values give more repeatable output. It does not make answers more accurate.
Token: The unit a model reads and generates, often a word or part of one. Cost and context limits are both counted in tokens.
Tool use: An agent's ability to call external systems such as APIs, databases or code execution in order to retrieve data or take action.
Trajectory: The whole path an agent took through a task — what was asked, what was retrieved, which tools were called, what came back, what it did next. Agents fail along the trajectory rather than at the final answer, so the trajectory is what has to be recorded.
Transformer: The neural network architecture behind modern language models. It computes relationships across its whole context in parallel, while still generating text one token at a time.
Vector database: A database that stores embeddings and supports fast search by meaning rather than by exact wording.
Zero-shot, one-shot, few-shot: Prompting with no examples, one example, or several. More examples fix a pattern more firmly, at the cost of context and money on every call.