Collaborator
The inflection point. The agent touches your actual artifacts - files, code, git. This is where conversation ends and execution begins.
Data Access
User → Agent → Code / Files / Artifacts
AAL 3 is where the agent stops talking and starts doing. It can read your files, write your code, run terminal commands, and commit to git. Every level below this is advisory. At AAL 3 and above, the agent is operational.
What This Level Is
Collaborator mode gives the agent direct access to your artifacts. It reads files, edits code, runs shell commands, and manages your git history. The agent is no longer a suggestion engine - it is a participant in your codebase.
This is also where most structural mismatches occur. Builders at APL 2 (in-editor assistants) often unknowingly grant AAL 3 access without the governance skills to match.
Tools at This Level
- Claude Code - Anthropic's CLI agent with full file and terminal access
- Cursor (Agent/Composer mode) - The same editor used at AAL2 in suggestions mode; in Agent mode it writes files and runs terminal commands directly
- Cline (agent mode) - VS Code extension in autonomous mode; routes to any model via OpenRouter, making DeepSeek via Cline, Minimax via Cline, etc. all AAL3
- Aider - Terminal-based agent that commits directly to git
- Gemini CLI - Google's command-line agent
- Codex CLI - OpenAI's terminal-based coding agent
- Ollama - Local model runner (Meta Llama, DeepSeek local, etc.) powering offline agents with file access
- Windsurf - In-editor agent with autonomous multi-file edit capability
What You Can Do
- Let the agent read and write files across your entire project
- Run shell commands, install packages, and start servers through the agent
- Have the agent commit changes directly to your git repository
- Delegate multi-file refactors and feature implementations
- Use the agent to run and fix tests autonomously
What You Cannot Do
- Assume the agent understands your full architecture from file access alone
- Safely skip reviewing diffs - silent changes are possible
- Operate without version control - you need a way to undo
- Coordinate multiple agents at this level - that is AAL 4
Risks at AAL 3
- Silent modifications: The agent can change files you did not ask it to - review every diff
- Git history pollution: Poorly governed commits are hard to untangle later
- Permission scope: Terminal access means the agent can reach anything your user account can - limit it where possible
- Context drift: Long agent sessions accumulate assumptions that diverge from your actual intent
Governance Essentials
Before You Operate at AAL 3
- Always commit before starting an agent session - you need a clean rollback point
- Review every diff before accepting - understand what changed and why
- Set explicit scope in your instructions - tell the agent which files are in and out of bounds
- Use a CLAUDE.md or project instructions file to give the agent stable context
When to Move Up
If your work requires multiple agents coordinating through defined workflows and schemas, you are looking at AAL 4: Operator. Most builders do not need it. AAL 3 with good governance covers the vast majority of use cases.