AI agent glossary

Agent Memory

What an agent stores and retrieves across steps or sessions — the difference between an agent that repeats itself and one that carries context forward.

AI Agent

A system that interprets an objective, decides its own steps, uses tools, takes action, and observes results — inside permissions and governance a person sets.

AI Assistant

A model wrapped in an interface for a human. It can help a person think, draft, summarise, or answer — but it doesn't work unattended or complete a process on its own.

AI Automation

A predefined sequence of steps, some of which may call a model. It executes known steps reliably and cheaply at scale, but can't handle a situation its designer didn't anticipate.

Chain of Thought

Producing intermediate reasoning before answering. Genuinely helpful for non-reasoning models on multi-step tasks; research shows it adds cost with little or no accuracy gain on dedicated reasoning models.

Context Engineering

Deciding what information reaches a model at each step — not just how a prompt is worded. Reliable agent behaviour depends on what the model can see, not on clever phrasing.

Context Rot

Degradation of AI accuracy as context grows, well before any technical size limit is reached — the reason the smallest set of high-signal information beats the largest set of possibly-relevant information.

Evaluation

Measuring whether a non-deterministic system is actually performing correctly, rather than assuming a good demo means reliable production behaviour.

Few-Shot Prompting

Including examples of the input/output pattern you want in a prompt, to guide behaviour without retraining. Two or three diverse, canonical examples typically outperform many similar ones.

Governance

The permissions, approval steps, and oversight structure that limit what an agent is allowed to do — the primary safety control for agentic systems.

Grounding

Basing AI output on real, supplied material rather than generalised training — the primary defence against hallucination.

Hallucination

Confident model output that isn't grounded in the facts or source material it was given — a core reason agents are scoped to answer only from provided knowledge.

Holistic Agent System

A governed system combining models, context, knowledge, memory, reasoning, planning, tools, actions, observation, evaluation, humans, other agents, and business constraints — able to own an outcome, be measured, be audited, and be improved.

Human-in-the-Loop

A design pattern requiring a human to approve, review, or intervene at a defined point before or after an agent acts — the dominant pattern in production agent systems today.

In-Context Learning

A model's ability to perform a new task from examples or instructions given in the prompt itself, without retraining — the mechanism behind few-shot prompting.

Multi-Agent System

Several agents coordinating toward a larger objective — parallelising, specialising, delegating, and reviewing each other's work.

Orchestration

Coordinating steps, tools, or multiple agents toward one outcome — deciding what runs when, and who hands off to whom.

Prompt Engineering

The deliberate design of instructions and inputs that guide an AI system toward a desired outcome — not a search for one magic phrase, but stating the objective, task, context, and constraints clearly.

Prompt Injection

Malicious or misleading instructions hidden in content an AI reads, attempting to redirect it from the operator's actual intent. No complete fix exists at the wording level — the defence is architectural.

Retrieval-Augmented Generation

Fetching relevant material and placing it in an AI's context so it answers from real sources rather than generalised training, ideally with citations.

Structured Output

Constraining an AI's generation to a defined schema, converting a request for a format into a guarantee of one — eliminating a whole class of downstream parsing failures.

Tool Use

An agent calling an external function or API to take an action outside of the conversation itself — booking a slot, sending an email, updating a record.

Verification

Independently checking a claim or result before relying on it — the highest-return reliability practice available, and distinct from asking the AI to check itself.