Agent Power Levels
APL measures your degree of control and execution capability. It's the 'what can you do' axis.
What are the 4 Agent Power Levels?
The 4 Agent Power Levels (APL) are: APL 1 Zero Setup (browser chat), APL 2 Light Setup (in-editor AI), APL 3 Dev Setup (CLI + codebase-wide operations), and APL 4 System Setup (multi-agent orchestration). Each represents a fundamentally different interaction model with your AI agent. The jump from APL 2 to APL 3 is where most people either level up or get exposed - it is the inflection point where the agent goes from suggesting to acting.
How does the APL framework work?
Agent Power Level is not about skill - it is about setup. Someone using ChatGPT in a browser (APL1) has a fundamentally different relationship with their agent than someone running Claude Code from a terminal (APL3), regardless of how experienced they are.
It also applies beyond software development. An author using AI to draft and edit manuscripts, a marketer automating content workflows, a manager generating reports from data, a designer iterating on copy with an AI assistant - all of these are APL users. The tools change; the levels do not. The question is always: how much can your setup do without you manually driving every step?
"Chatting in a browser. Nothing installed."
- Ideas and brainstorming
- Draft code snippets
- Concepts and explanations
- Simple prototypes
- No real project access
- No execution
- No tools
- Manual copy/paste only
ChatGPT, Claude.ai, Gemini Chat, Grok, DeepSeek Chat, z.ai, Lovable
"You install an IDE with built-in agent."
- File editing
- Repo awareness
- Real code changes
- Basic refactoring
- No terminal access
- No system commands
- Depends on local skills
Cursor (suggestions), Windsurf, VS Code + Copilot, Cline (suggestions)
"Your agent works inside your editor and your terminal."
- Runs commands
- Fixes errors
- Installs packages
- Starts servers
- Handles real dev workflows
- More setup required
- Requires trust + supervision
Claude Code, Cursor Agent, Aider, Gemini CLI, Codex CLI, Ollama, DeepSeek via Cline, OpenRouter
"Multiple agents + tools, running in autopilot mode."
- Tool calling
- Multiple agents working together
- Workflow orchestration
- Automated builds, tests, deployments
- Requires setup knowledge
- Most powerful - most responsibility
LangGraph, CrewAI, AutoGen, AWS Bedrock, Azure OpenAI, Vertex AI
APL comparison table
| Level | Name | Mode | Key capability | Example tools | Risk factor |
|---|---|---|---|---|---|
| APL 1 · Zero Setup | Zero Setup | UI-Only | Ideas, drafts, explanations | ChatGPT, Claude.ai, Gemini | Low - no execution |
| APL 2 · Light Setup | Light Setup | IDE Power | File editing, repo awareness | Cursor, Windsurf, Copilot | Low - suggestions only |
| APL 3 · Dev Setup | Dev Setup | IDE + CLI | Commands, packages, servers | Claude Code, Aider, Gemini CLI | Medium - agent acts |
| APL 4 · System Setup | System Setup | Ecosystem | Multi-agent orchestration | LangGraph, CrewAI, AutoGen | High - autonomous |
Why is the same model at different APL levels?
The same underlying AI model can appear at multiple APL levels depending on how you access it. DeepSeek Chat is APL1 - you open a browser and talk to it. DeepSeek via Cline is APL3 - the same model, now powering an agent that reads and writes your codebase. The level is determined by the access method, not the model name.
The same applies to tools with multiple modes: Cursor is APL2 in suggestions mode and APL3 in Agent/Composer mode. GitHub Copilot is APL2 in the VS Code chat interface and APL3 when used via API in a pipeline. When in doubt, ask: is the agent acting, or am I still approving every move?
What happens at the APL 2 to APL 3 jump?
The biggest shift in the entire framework happens between APL2 and APL3. At APL2, your AI assistant suggests - you review and approve every change. At APL3, your agent acts - it reads, writes, and runs commands across your entire project on its own. This is the inflection point where most mismatches occur - and where the AI Setup Snapshot becomes most valuable.