The AI glossary

Agent evals are repeatable tests that measure whether an AI agent actually completes tasks correctly — scoring real runs against expected outcomes rather than judging output by eye.

An agent loop is the cycle an AI agent repeats to make progress — decide the next step, call a tool, observe the result, update its state — running until the goal is met or a limit stops it.

Agent memory is the mechanism that lets an AI agent carry information across steps and sessions — short-term context within a run, and long-term stored facts it can retrieve later.

An agentic workflow is a multi-step process driven by an AI agent that plans, uses tools, and adapts — instead of following a fixed, pre-scripted sequence.

An AI agent is a system that pursues a goal by reasoning, choosing tools, and taking multi-step actions — not just answering a single prompt.

An AI Operating System is a workspace where autonomous AI agents, a mesh of tools, and native apps work together as one environment — the way a traditional OS unifies programs and hardware.

An API connector is the integration layer that wraps a third-party service's API — handling auth, request shaping, pagination, retries and errors — so an agent can call it as a simple tool.

Approval tiers classify an agent's actions by how much damage they can do — typically read, write and destructive — and require human sign-off only on the tiers where a mistake actually matters.

Automation runs a defined task or workflow automatically on a trigger or schedule, without manual effort each time.

An autonomous agent is an AI system that pursues a goal without step-by-step human direction — deciding its own next actions, recovering from failures, and stopping when the objective is met.

A capability mesh treats every tool, connector and action as searchable data, so an agent can discover and invoke the right capability at runtime instead of being hard-wired to a fixed set.

Chunking is splitting long documents into smaller passages before embedding them, so retrieval can return the specific paragraph that answers a question rather than a whole 200-page manual.

A connector is the secure link between an AI system and an external tool (like Gmail or GitHub) that lets an agent read from and act on that service.

Context engineering is the practice of deciding what goes into a model's context window on each call — instructions, retrieved data, tool results, history — so the model has what it needs and little else.

The copilot-vs-agent distinction is about who does the work: a copilot suggests inside a tool you are actively using, while an agent takes the goal and performs the steps itself.

A digital worker is an AI agent framed as a member of staff rather than a feature — it owns an ongoing role, works across the same tools a person would, and is judged on completed outcomes.

An embedding is a list of numbers representing a piece of text, image or audio in a way that places similar meanings close together — the numeric form that makes semantic search possible.

Fine-tuning is further training an existing model on your own examples to adapt its behaviour — teaching it a format, tone or task pattern that prompting alone can't reliably produce.

Function calling is the model capability behind tool use: given machine-readable function schemas, the model emits a structured call with typed arguments instead of prose, and your code runs it.

Guardrails are the constraints around an AI system — input filters, output checks, permission limits and approval gates — that keep it inside intended behaviour regardless of what the model decides.

A hallucination is when an AI model produces confident but false or fabricated information.

Human-in-the-loop (HITL) is a design where an AI agent pauses for human review and approval before taking consequential or irreversible actions.

Idempotency is the property that performing an operation more than once has the same effect as performing it once — so a retried request can't accidentally charge, send or create twice.

Inference is the act of running a trained model to produce output — the request-time work of turning a prompt into a response, as opposed to the training that created the model.

A large language model (LLM) is an AI trained on vast text to predict and generate language, powering chat, reasoning and tool-use in modern AI systems.

Latency is the delay between a request and its response — for language models, usually split into time-to-first-token and the rate tokens stream after it, and for agents, the sum across every step.

The Model Context Protocol (MCP) is an open standard for connecting AI models to external tools and data sources through a common interface.

Model routing is automatically sending each task to the most suitable AI model — balancing quality, speed and cost across multiple providers.

A multi-agent system splits work across several specialised AI agents that coordinate — often a supervisor delegating to sub-agents — instead of asking one agent to do everything alone.

OAuth scopes are the specific permissions an app requests when you connect an account — like read-only calendar access — defining exactly what it can do rather than granting your whole account.

Orchestration coordinates multiple models, agents and tools so they work together toward a goal, handling sequencing, state and hand-offs.

Planning is the step where an AI agent turns a goal into an ordered set of sub-tasks before acting, so it can sequence tool calls deliberately rather than improvising one step at a time.

Prompt injection is an attack where malicious instructions hidden in content trick an AI agent into doing something it shouldn't.

Rate limiting caps how many requests a client may make in a window — say 100 per minute — protecting a service from overload and returning 429 Too Many Requests when the cap is exceeded.

ReAct is an agent pattern where the model alternates reasoning and acting — thinking a step through, calling a tool, reading the result, then reasoning again — until the task is done.

Retrieval-augmented generation (RAG) improves an AI's answers by first retrieving relevant documents and feeding them to the model as context.

A scoped token grants an agent only the specific, time-limited permissions it needs for one run — limiting the blast radius if something goes wrong.

Semantic search retrieves results by meaning rather than keyword overlap, so a query like how do I stop paying finds a page titled cancel your subscription with no words in common.

Structured output is making a model return data in a guaranteed shape — typically JSON matching a schema — so downstream code can parse it reliably instead of scraping prose.

A system prompt is the standing instruction given to a model ahead of the conversation — its role, rules, tone and constraints — that shapes every response in the session.

Temperature is a sampling setting that controls how random a model's output is — near 0 it picks the most likely next token almost every time, and higher values let unlikelier ones through.

A token is the unit a language model reads and writes — roughly a word-piece of about four characters in English — and the unit that context limits, pricing and speed are all measured in.

Tool calling is when an AI model invokes external functions or APIs to fetch data or take actions, rather than only generating text.

A trigger is the event or condition that starts an automated run — a schedule firing, a webhook arriving, a file landing, a label being applied — without anyone pressing a button.

Vector search finds results by meaning rather than exact keywords, by comparing numerical embeddings of text.

A webhook is a reverse API call: instead of you polling a service for changes, the service sends an HTTP POST to a URL you registered the moment an event happens.

Workflow automation is stringing the steps of a business process together across systems so it runs on its own — a trigger fires, data moves, conditions branch, and each step happens without a person.