Plain-English definitions for every AI coding assistant term used on this site — from agents and models to MCP and instruction files.

This is a lookup tool: each entry gives you the one-paragraph definition and links to the full concept page where one exists. Use it when you encounter a term you don’t recognise, or to check the precise meaning used on this site.

Definitions reflect the meaning of each term as used across this site. Where a term has multiple common meanings in the industry, we note the distinction.

A

AICA (AI Coding Assistant)

An AI-powered product that helps you write, understand, and review code — built on top of one or more large language models. GitHub Copilot, Cursor, and Windsurf are all AICAs. See What is an AI Coding Assistant?

AICA Onion

The layered mental model this site uses to organise AI coding assistant concepts, from the raw model at the centre to orchestrators and tools at the outer layers. Each layer adds capability on top of the one below. See the Start Here page for the full diagram.

Agent

An AI system that pursues a goal autonomously — calling tools, reading files, writing code — without requiring step-by-step human approval. Agents loop, self-correct, and act; chat mode does not. See What is an Agent?

Agent Mode

A capability inside GitHub Copilot that lets it autonomously edit your code, determine which files to change, and run terminal commands — all within your local IDE. Agent mode is synchronous and local; it is distinct from the Copilot coding agent which runs in the cloud. See Copilot Agent Mode

AGENTS.md

A Markdown instruction file that tells AI agents how to behave in a repository — where to run tests, which commands to use, and what conventions to follow. The nearest AGENTS.md up the directory tree takes precedence. This is a community-driven convention (not a GitHub-owned spec) and is also accepted as CLAUDE.md or GEMINI.md at the repo root. See Copilot Instructions

Autonomous agent

An agent that completes a multi-step goal without asking for human approval at each decision point. The Copilot coding agent and third-party agents like Anthropic Claude coding agent are examples. See What is an Agent?

C

Chat mode

The standard request-response interaction with an AI coding assistant — you ask a question, it answers, nothing happens autonomously. Chat mode does not loop, call tools, or make file edits on its own. It is the opposite of agent mode.

Claude (Anthropic)

Anthropic’s family of large language models — including Claude Haiku, Claude Sonnet, and Claude Opus at various versions. Claude models are available as selectable models inside GitHub Copilot. Note: the Claude model is not the same as the Anthropic Claude coding agent, which is a separate autonomous agent product. See Models vs Agents

CLAUDE.md

An alternative name for an agent instruction file, placed at the repository root and used by Anthropic-backed agents as their instruction source. Functionally equivalent to AGENTS.md for agents that recognise it. See Copilot Instructions

Coding agent (Copilot)

GitHub Copilot’s cloud-based autonomous agent — assign it a GitHub issue and it works asynchronously in a GitHub Actions environment, writes code, and opens a pull request for you to review. This is entirely distinct from agent mode, which runs locally in your IDE. See Copilot Coding Agent

Completion / autocomplete

The original AI coding assistant feature: inline code suggestions that appear as you type, which you accept or dismiss. This is a single-turn, non-autonomous interaction — the model predicts what you intend to type next.

Context window

The total amount of text (measured in tokens) that a model can read and process in a single request. Everything the model “knows” during a response — your code, chat history, instruction files, tool outputs — must fit within the context window.

Copilot Instructions (.github/copilot-instructions.md)

A Markdown file that provides repository-wide custom instructions for every GitHub Copilot request in that repo. It is injected into the model’s prompt automatically — it is prompt context, not an agent. See Copilot Instructions

Cursor

An AI coding assistant built on VS Code that uses large language models to provide inline completions, chat, and agent-like editing capabilities. Cursor supports MCP and uses .cursorrules for repository-specific instructions. See The AI Coding Assistant Landscape

Cursor rules (.cursorrules)

A repository-level instruction file used by Cursor to customise its behaviour in a specific project — analogous to .github/copilot-instructions.md in GitHub Copilot. It is injected as prompt context and is not an agent.

D

Devin

An autonomous coding agent by Cognition Labs, designed to complete entire software engineering tasks end-to-end with minimal human involvement. Devin operates at Layer 2 (agent) of the AICA Onion and is an example of a stand-alone coding agent product rather than a feature within an existing assistant.

F

Foundation model

A large model trained on broad data at enormous scale, used as the base for many downstream applications. GPT-5.4, Claude Sonnet 4.6, and Gemini 2.5 Pro are all foundation models. See What is a Model?

G

Gemini (Google)

Google’s family of large language models, including Gemini 2.5 Pro and the Gemini 3 Flash and Pro variants. Gemini models are available as selectable models inside GitHub Copilot. See Copilot Models

GitHub Copilot

GitHub’s AI coding assistant — a product that wraps large language models to provide completions, chat, agent mode, and cloud-based coding agents across IDEs, the GitHub website, and the CLI. Copilot is unusual in the AICA landscape because it operates simultaneously as a Layer 1 assistant and a Layer 3 orchestrator. See GitHub Copilot: An Overview

GPT (OpenAI)

OpenAI’s family of large language models — including GPT-4.1, GPT-5, and the GPT-5.x Codex variants. GPT models are available as selectable models inside GitHub Copilot. See Copilot Models

Grok (xAI)

xAI’s family of large language models, including Grok Code Fast. Grok models are available as selectable models inside GitHub Copilot. See Copilot Models

H

Hallucination

When a model generates a response that is confidently stated but factually incorrect or entirely fabricated. Hallucinations occur because models predict likely token sequences rather than retrieving verified facts. Always verify AI-generated code and references.

I

Inference

The act of running a trained model to generate output from an input prompt. Every time you send a message to an AI coding assistant, you are triggering inference. Inference consumes compute and (in most products) is metered as tokens processed.

Instruction file

A generic term for any Markdown file that shapes how an AI agent or assistant behaves in a repository — including copilot-instructions.md, AGENTS.md, CLAUDE.md, and .cursorrules. Instruction files are prompt context injected before the model responds; they are not agents. See Copilot Instructions

L

Large Language Model (LLM)

A statistical model trained on large amounts of text that generates token sequences in response to a prompt. An LLM is the engine at the centre of every AI coding assistant; it is not itself a product, an agent, or GitHub Copilot. See What is a Model?

M

MCP (Model Context Protocol)

An open standard that defines how AI models and agents connect to external tools, data sources, and services — described officially as “USB-C for AI.” MCP is a protocol, not an agent or a model. See What is MCP? and Extensions, Tools & MCP

MCP server

A service that exposes callable tools following the MCP standard — for example a web search tool, a database query tool, or a GitHub API tool. The MCP server is not an agent; the agent calls the server’s tools when it decides to. See Extensions, Tools & MCP

Model

The underlying AI system — trained on data, running on GPUs — that generates text in response to a prompt. The model is the engine; the AI coding assistant is the product wrapper built around it. You can often swap models within the same assistant. See What is a Model?

Model provider

A company that trains and serves large language models — for example OpenAI (GPT), Anthropic (Claude), Google (Gemini), and xAI (Grok). Model providers are distinct from AI coding assistant products; GitHub Copilot lets you choose models from multiple providers. See The AI Coding Assistant Landscape

O

Orchestrator

A system that coordinates multiple agents or models, routing work between them and managing the overall task flow. GitHub Copilot acts as an orchestrator when it delegates tasks to the Copilot coding agent or third-party agents like Anthropic Claude. See GitHub Copilot: An Overview

P

Prompt

The input you send to a model or AI assistant — your question, instruction, or code snippet. Everything the model generates is a response to the prompt it received.

Prompt context

Everything the model can see in its context window when generating a response — your message, chat history, instruction files, retrieved code snippets, and tool outputs. Shaping prompt context well is how instruction files and MCP servers influence model behaviour.

Public preview

A feature that has been released for broad testing but is not yet generally available. Features in public preview may change, be removed, or have limited availability. Third-party agents in GitHub Copilot (Anthropic Claude, OpenAI Codex) were in public preview as of early 2026.

R

Reasoning model

A large language model that performs extended internal reasoning — often called chain-of-thought — before producing its final answer. Reasoning models tend to be slower and more expensive per request but handle complex multi-step problems more reliably. OpenAI’s o-series and Anthropic’s extended-thinking variants are examples.

S

Skill (MCP tool)

An individual callable function exposed by an MCP server — for example search_web, query_database, or create_issue. An agent decides when and whether to call a skill; the skill itself is just a function definition. See Skills and Tools and Extensions, Tools & MCP

System prompt

Hidden instructions provided to a model before any user interaction, used by the product or assistant to set behaviour, persona, and constraints. You typically do not see the system prompt; it is injected by the assistant (e.g. GitHub Copilot) before your message reaches the model.

T

Third-party agent

An autonomous coding agent from a provider other than GitHub, accessible through GitHub Copilot’s orchestration infrastructure. As of early 2026, Anthropic Claude coding agent and OpenAI Codex agent are available as third-party agents within Copilot (public preview). See What is an Agent?

Token

The basic unit of text that a model processes — roughly three-quarters of a word on average for English text. Models are billed by tokens consumed, and their context windows are measured in tokens. A token is not a word; punctuation and spaces are also tokenised.

Tool (in MCP context)

A callable function that an agent can invoke at runtime to interact with an external system — such as reading a file, querying an API, or running a terminal command. Tools are the mechanism by which agents act on the world beyond the context window. See Skills and Tools

W

Windsurf

An AI coding assistant by Codeium that supports chat, completions, and agent-like autonomous editing. Windsurf supports MCP, making its agents extensible with the same tool servers used by other MCP-compatible assistants. See The AI Coding Assistant Landscape

Dive deeper

📍 Glossary · All Layers · Back to the AICA Onion

What’s next?

Updated: