GitHub Copilot: An Overview
GitHub Copilot is not just one thing — it’s an AI coding assistant (Layer 1) that can also act as an orchestrator (Layer 3), routing work to models, its own agents, and even third-party agents.
The 30-second version: Copilot started as an autocomplete tool. Today it’s a multi-layer product: it lets you choose models, enables autonomous agent mode, runs a cloud-based coding agent, and can route to third-party agents. It’s the most complete example of the AICA Onion in one product.
Copilot across the layers
Copilot is unusual because it doesn’t sit at just one layer of the AICA Onion. It spans from the model selection at Layer 0 all the way up to orchestration at Layer 3.
| Layer | What Copilot does here |
|---|---|
| Layer 0 — Model | Copilot lets you choose which model powers your responses (GPT-5.4, Claude Sonnet 4.6, Gemini 2.5 Pro, and others) |
| Layer 1 — AI Coding Assistant | Copilot IS the AI coding assistant — the IDE extension, chat interface, and context manager |
| Layer 2 — Agent | Copilot has two agent products: Agent Mode (local, sync) and the Copilot Coding Agent (cloud, async) |
| Layer 3 — Orchestrator | Copilot can route to third-party agents (Claude coding agent, OpenAI Codex agent) |
| Layer 4a — Instructions | Copilot reads copilot-instructions.md and AGENTS.md to customise behaviour |
| Layer 4b — Tools/MCP | Copilot supports MCP servers to extend what agents can do |
The two things people most often confuse
⚠️ Two confusions to avoid:
- Copilot’s model and Copilot itself are different things. Changing your model (GPT-5.4 vs Claude) doesn’t change Copilot — it changes the engine Copilot calls.
- Agent Mode and the Copilot Coding Agent are two different products with the same word in their names. Agent Mode works in your IDE. The Copilot Coding Agent works in GitHub Actions and creates PRs.
What’s in the Copilot section
Each sub-page goes deep on one part of Copilot. Start with whichever layer you’re most curious about — or work through them in order.
- Models in Copilot — which models are available, how to switch
- Agent Mode — autonomous multi-file editing in your IDE
- Copilot Coding Agent — async cloud agent that creates PRs
- Extensions & MCP — third-party tools and integrations
- Instruction Files —
copilot-instructions.mdandAGENTS.md - Enterprise & Trust Features — content exclusions, data residency, policy controls, audit logs, and IP indemnity for org deployments
Practical patterns (Layer 2 skills — how to use these features effectively):
- Practical Copilot Patterns — opinionated guidance on choosing the right mode, prompting strategy, and when to skip AI entirely
- Prompting Agent Mode — what makes agent mode prompts different and how to write them
- Writing Effective Copilot Instructions — practical patterns for
copilot-instructions.mdfiles - When NOT to Use Copilot — security, learning, compliance, and judgment require human reasoning
Go deeper
If you want the full conceptual picture before diving into sub-pages, start with the AICA Onion layer map or read Models vs Agents — that page is the single best explainer for why the Copilot vocabulary gets confusing.
Related Concepts
- What is an AI Coding Assistant? — The Layer 1 concept that Copilot is the primary example of
- Models vs Agents — The core distinction that explains why Copilot’s vocabulary is confusing
- What is an Agent? — The Layer 2 concept behind Copilot Agent Mode and the Copilot Coding Agent
📍 Layers 1–3 (highest: Layer 3) — GitHub Copilot spans multiple layers · Back to the AICA Onion
What’s next?
- Agent Mode — how Copilot becomes an autonomous agent inside your IDE