Start Here: Understanding AI Coding Assistants
AI coding assistants look complicated because nobody explains how the layers fit together. This page does.
AI coding assistant tooling comes with a lot of overlapping vocabulary — “model,” “agent,” “Copilot,” “instructions,” “MCP” — and very little clarity about what those words actually mean. This site cuts through that. Every page anchors to one specific layer of the stack.
The Confusion This Site Solves
When you switch the model in GitHub Copilot to “Claude,” are you using Anthropic’s model, Anthropic’s agent, or something else entirely? The answer is: it depends, and that ambiguity is the problem.
The vocabulary of AI coding assistants has not kept pace with the tools themselves. Words get reused across different layers of the stack with different meanings. This site maps those layers clearly, so you can stop guessing.
The AICA Onion
Think of AI coding assistants as a stack of layers — each one built on the one below it. We call this the AICA Onion.
| Layer | Name | What it is | Examples |
|---|---|---|---|
| 0 | The Model | The math engine — predicts text given a prompt | GPT-5.4, Claude Sonnet 4.6, Gemini 2.5 Pro |
| 1 | The AI Coding Assistant | The product you use — responding to you: autocomplete, chat, inline suggestions | GitHub Copilot, Cursor, Windsurf |
| 2 | The Agent | Same product, now acting — loops, uses tools, reads files, makes decisions autonomously | Copilot agent mode, Copilot coding agent |
| 3 | The Orchestrator | Same product, now routing — delegates your request to a third-party agent instead of handling it directly | GitHub Copilot (Agents tab) |
| 4a | Instructions | Markdown files that shape behaviour — not agents | copilot-instructions.md, AGENTS.md |
| 4b | Tools / MCP | Callable external capabilities — not agents | MCP servers, Copilot Extensions |
Each layer depends on the one below it. The model does nothing without a product to call it. An agent does nothing without a model to reason with.
Layers 1, 2, and 3 can all be the same product — GitHub Copilot. The difference is what it’s doing. At Layer 1 it’s responding to you; at Layer 2 it’s acting on your behalf; at Layer 3 it’s routing your request to a third-party agent. When you “switch to Claude” in Copilot, you could be at Layer 0 (swapping the model) or Layer 3 (routing to Claude’s Agent) — the answer depends on which option you picked.
⚠️ “Agent” means two different things in GitHub Copilot — Agent Mode is a local IDE feature for multi-file edits; the Copilot Coding Agent is a separate cloud-based product that works asynchronously via GitHub Actions. The Agent Mode and Copilot Coding Agent pages explain the difference.
Where to Start
Read these in order — each page builds on the last:
- What is a Model? — Layer 0. The foundation everything else is built on.
- What is an AI Coding Assistant? — Layer 1. The product you actually use day-to-day.
- What is an Agent? — Layer 2. When the assistant acts autonomously on your behalf.
- Models vs Agents — The most common confusion, addressed head-on.
Already know the basics? Jump straight to Models vs Agents — it’s the highest-value page on the site.
Begin reading
- What is a Model? — start at Layer 0
- What is an AI Coding Assistant? — Layer 1
- GitHub Copilot Overview — jump straight to Copilot
📍 Start Here · All Layers · Documentation Index