AI Glossary 2026: Essential Terms in AI, Agentic Workflows, Quantum & Robotics — With Real Workflow Examples

An actuator is the part of a robot that creates movement — a motor, servo, or piston that turns a control signal into physical action like turning a wheel or closing a gripper. It's the muscle to the robot's brain.

An AI agent is a system that uses a language model to decide what to do, take actions like searching the web or calling tools, look at the results, and keep going until a goal is met — rather than just answering once.

An API key is a secret password that lets your app talk to a service like an AI model provider. Whoever holds the key can run requests and rack up charges, so it must be kept private.

Chain-of-thought is when a model works through a problem step by step, in writing, before giving its final answer. Showing its working tends to make it more accurate on reasoning, maths, and logic tasks.

The context window is the maximum amount of text — your prompt plus the model's reply — that a model can consider at once. Think of it as the model's short-term memory: anything beyond it is forgotten or must be re-supplied.

A cron or schedule runs a task automatically at set times — every hour, every weekday morning, or once a month. It's how you make automations happen on a clock instead of waiting for an event.

A digital twin is a live virtual copy of a real machine, robot, or system, kept in sync with sensor data. You can test changes, predict problems, and simulate scenarios on the twin before touching the real thing.

Distillation trains a smaller, faster model to mimic a larger, more capable one. You get most of the quality at a fraction of the size and cost.

Edge AI runs models directly on a device — a phone, camera, sensor, or robot — instead of in the cloud. It responds instantly and works offline, which matters when there's no time or connection to call a server.

An embedding turns text (or an image) into a list of numbers that captures its meaning. Pieces of content with similar meaning end up with similar numbers, which is how computers find related items.

Entanglement links two or more qubits so that their states are tied together — measuring one instantly tells you about the others, no matter the distance. It's a key resource that makes quantum computing powerful.

An eval is a repeatable test that measures how well your AI system performs on a set of example cases. It turns 'this feels better' into a number you can track as you change prompts, models, or settings.

Fan-out is splitting one job into many parallel pieces — for example processing a list of 500 records as 500 simultaneous tasks — so the work finishes far faster than doing them one by one.

Few-shot prompting means showing a model a handful of examples of the task before asking it to do the real one. The examples teach it the pattern you want without any retraining.

Fine-tuning means further training an existing model on your own examples so it gets better at a specific task or adopts a particular style. It changes the model itself, unlike prompting.

A foundation model is a large, general-purpose AI trained on huge amounts of data that can then be adapted to many tasks. The big chat and coding models you use are foundation models.

A GPU is the chip that does the heavy maths for AI, and VRAM is its dedicated memory. The amount of VRAM largely decides how big a model you can run, especially when hosting models yourself.

Guardrails are the safety checks around an AI system that keep it from doing or saying the wrong thing — blocking harmful content, off-topic answers, or risky actions before they reach the user.

A hallucination is when a model states something false but sounds confident — inventing a fact, a citation, or a feature that doesn't exist. It happens because models predict plausible text, not verified truth.

An idempotent action gives the same result whether it runs once or many times. It matters in automation because retries and duplicate triggers shouldn't, say, charge a customer twice or send the same email five times.

Inference is the act of running a trained model to get an answer — every time you send a prompt and get a reply, that's inference. It's separate from training, which is how the model was built.

The KV cache is the model's memory of the tokens it has already processed, so it doesn't redo that work for every new word it generates. It speeds up replies but takes up memory that grows with context length.

Latency is how long one request takes to answer; throughput is how many requests you can handle per second. Tuning for one often costs the other, so you pick based on whether speed-per-user or total volume matters more.

A local LLM is a language model that runs on your own computer or server instead of a cloud service. Your data never leaves your machine, but you need enough hardware to run it.

LoRA is a cheap way to fine-tune a model: instead of retraining the whole thing, it trains a small add-on that nudges the model's behaviour. The add-ons are tiny and easy to swap.

MCP is an open standard, introduced by Anthropic, that lets AI apps connect to tools and data sources through one common interface — like a USB port for AI. Build a connector once and any MCP-aware app can use it.

Mixture-of-experts is a model design where only a few specialised sub-networks ('experts') activate for each input, instead of the whole model. This makes very large models faster and cheaper to run.

Multi-agent orchestration is when several specialised AI agents work together on a task — for example one plans, one researches, and one writes — coordinated so their outputs combine into a finished result.

Post-quantum cryptography is a new generation of encryption designed to stay secure even against future quantum computers, which could break much of today's encryption. Organisations are starting to adopt it now.

Prompt engineering is the craft of writing clear instructions to a language model so it gives you the result you want. Small changes in wording, examples, or structure can dramatically improve the output.

Quantization shrinks a model by storing its numbers with less precision, so it uses less memory and runs faster — usually with only a small drop in quality. It's how big models fit on smaller hardware.

Quantum annealing is a specialised kind of quantum computing built to find good solutions to optimisation problems — like the cheapest route or best schedule — by letting the system settle into a low-energy state.

A qubit is the basic unit of a quantum computer, like a bit in a normal computer — but instead of being just 0 or 1, it can be a blend of both at once. This is what gives quantum machines their unusual power.

RAG is a technique that fetches relevant documents from your own data and feeds them to the model before it answers, so the reply is grounded in your content instead of just the model's general knowledge.

A rate limit is the cap on how many requests you can make to a service in a given time. Go over it and the service temporarily refuses your calls, so high-volume automations must pace themselves.

ReAct is a pattern where a model alternates between thinking and acting: it reasons about what to do, takes an action, sees the result, then reasons again. This loop lets it work through multi-step problems instead of guessing in one shot.

Reinforcement learning teaches an agent by trial and error: it tries actions, gets rewards or penalties, and gradually learns the behaviour that earns the most reward. It's widely used to train robots and game-playing agents.

Sim-to-real is the practice of training a robot in a virtual simulation, then transferring what it learned to the physical world. Simulation is cheap, fast, and safe, but the real world always differs, which is the challenge.

SLAM is how a robot builds a map of an unknown space while figuring out where it is within that map at the same time. It's what lets a robot vacuum or drone navigate somewhere it has never been.

Speculative decoding speeds up a big model by having a small, fast model guess several next words ahead, which the big model then verifies in one go. Correct guesses are kept, so you get faster output with the same quality.

Superposition is a quantum effect where a qubit holds a combination of 0 and 1 at the same time, rather than settling on one. Measuring it forces a single outcome, but until then it explores possibilities together.

A system prompt is the standing set of instructions that tells a model who it is, how to behave, and what rules to follow throughout a conversation. The user's messages come after it, but the system prompt sets the ground rules.

Teleoperation is controlling a robot remotely while a human steers it. Beyond direct control, it's also a way to collect demonstration data that trains a robot to do the task on its own later.

Temperature is a dial that controls how random or creative a model's output is. Low temperature gives focused, predictable answers; high temperature gives more varied, surprising ones.

A token is a small chunk of text — often a word or part of a word — that a language model reads and writes in. Models count and bill in tokens rather than characters, so a rough rule of thumb is that one token is about four characters or three-quarters of an English word.

Tool use is how a model reaches outside itself to do things it can't do alone — look up live data, run code, send an email, or query a database. You describe the available tools and the model decides when to call them.

A vector database stores embeddings — the numeric meaning of your content — and finds the closest matches to a query fast. It's the search engine that powers semantic search and RAG.

A webhook is a way for one app to instantly notify another when something happens, by sending a small message to a URL. It's how automations react to events the moment they occur, instead of constantly checking.

Workflow automation chains together apps and steps so work happens without manual effort — when something triggers, a sequence of actions runs by itself. Adding AI lets those steps make decisions or write content.