Models vs Agents
A model generates text. An agent acts on the world. These are different layers — you can have one without the other.
A model is the engine: it predicts tokens. An agent is the vehicle: it uses an engine, plus tools, plus a decision loop, to complete tasks. You don’t interact with models directly — you interact with agents and assistants, which use models under the hood. The confusion happens because the same company (Anthropic, OpenAI) makes both the model and an agent product that shares the same name.
Are they the same thing?
⚠️ “But Copilot uses Claude — isn’t Claude an agent?” Claude is both a model family (Claude Sonnet 4.6, Claude Opus 4.6 — text prediction engines) and a separate product: the Claude coding agent (an agent product that uses Claude models). The name overlap is genuinely confusing. Always ask: are we talking about the model or the product?
Side by side
| Model | Agent | |
|---|---|---|
| What it does | Predicts the next token(s) | Takes actions to complete a goal |
| Awareness | No memory, no state, no tools | Has tools, can maintain state within a task |
| Autonomy | None — responds to a single prompt | Can plan and execute multi-step sequences |
| Examples | GPT-5.4, Claude Sonnet 4.6, Gemini 2.5 Pro | Copilot agent mode, Copilot coding agent, Claude coding agent |
| Runs where? | On a server, called via API | In an IDE, in GitHub Actions, or in the cloud |
| Can it edit your files? | No | Yes (with tool access) |
| Can you swap it? | Yes — Copilot lets you pick your model | Sometimes — Copilot hosts its own agents but also third-party ones |
In practice — how this shows up in GitHub Copilot
Scenario 1: Changing your model
In GitHub Copilot Chat, you can select GPT-5.4, Claude Sonnet 4.6, or Gemini 2.5 Pro. You’re choosing which model (Layer 0) powers your responses. The AI coding assistant wrapper — Copilot itself, Layer 1 — stays the same.
Scenario 2: Using agent mode
When you turn on agent mode in Copilot, you’re now talking to an agent (Layer 2) — not just a model. The agent can edit multiple files, run terminal commands, and loop until it’s done. It still uses a model under the hood, but you’re interacting with the agent layer.
Scenario 3: Third-party agents in Copilot
When you route a task to the Claude coding agent or OpenAI’s Codex agent from within Copilot, you’re asking Copilot — now acting as an orchestrator, Layer 3 — to hand off to a third-party agent product. That agent has its own model powering it. You’re not changing Copilot’s model; you’re routing to a different agent entirely.
The quick test
💡 Quick test: Can you swap it out for a different one doing the same job? If yes — it’s probably a model. Does it decide its own next steps? If yes — it’s behaving like an agent.
Dive deeper
Each layer has its own page if you want to go deeper on any of them:
- What is a Model? — Layer 0: the statistical engine
- What is an AI Coding Assistant? — Layer 1: the product wrapper
- What is an Agent? — Layer 2: autonomous task execution
- Copilot Agent Mode — the IDE-local agent experience
- Copilot Coding Agent — the cloud-based async agent that creates PRs
Related Concepts
- What is a Model? — The full explanation of Layer 0
- What is an Agent? — The full explanation of Layer 2
📍 Spans Layers 0–2 · highest: Layer 2 · Back to the AICA Onion
What’s next?
- Copilot Overview — see how GitHub Copilot spans all three layers simultaneously