Dev Setup
Codebase-wide agent operations. The agent reads, writes, executes, and iterates across your entire project.
Overview
APL3 is the inflection point. The agent goes from assistant to operator. It can modify files, run commands, and iterate on its own output. This is where leverage jumps - and so does risk.
What This Level Is
Dev Setup means the agent has access to your full codebase and can operate across it autonomously. It does not wait for you to point it at a file - it navigates, reads, edits, tests, and commits on its own.
This is the inflection point in the APL framework. Everything before this is assistance. Everything after is execution.
Tools at This Level
- Claude Code - Anthropic's agentic coding CLI
- Cursor (Agent/Composer mode) - The same editor as APL2, but in agent mode it runs terminal commands and edits autonomously across the full codebase
- Cline (agent mode) - VS Code extension in autonomous mode; can route to any model via OpenRouter - DeepSeek via Cline, Minimax via Cline, etc.
- Aider - Terminal-based agent that commits directly to git
- Gemini CLI - Google's command-line AI agent
- Codex CLI - OpenAI's terminal-based coding agent
- Ollama - Run local LLMs via terminal (Meta Llama, DeepSeek local, etc.) for private, offline code generation
- OpenRouter - API aggregator used as a model backend for CLI agents and Cline, enabling access to DeepSeek, Minimax, and others at APL3
What You Can Do
- Full codebase refactors across multiple files
- Automated test generation and execution
- Feature implementation from natural language descriptions
- Bug investigation across the entire project
- Documentation generation from code analysis
- Git operations - commits, branches, diffs
What You Cannot Do (Safely)
- Deploy to production without human review
- Trust all generated code without verification
- Let multiple agents coordinate without governance
- Skip testing because "the agent wrote it"
Risks at This Level
- Silent breakage: The agent edits 15 files but introduces a subtle bug in file 12
- Permission creep: Granting system access "just this once" becomes permanent
- Context overflow: The agent loses track of what it changed across a large session
- Overconfidence: The agent says "done" but the result does not match the intent
Safe Practices
- Review diffs before committing - always
- Use git branches for agent-driven changes
- Run tests after every agent session
- Set clear boundaries on what the agent can access in your CLAUDE.md
- Use the AI Testing Agents pattern
When to Level Up
If you need multiple agents working together, shared state, or orchestrated workflows, you are looking at APL4: System Setup. Most builders do not need APL4 - master APL3 first.