Multi-Agent Skills Development

An orchestration system that leverages Codex, Claude, Copilot, and Gemini CLIs for collaborative skills development. Supports multi-agent workflows with configurable agent assignments and A/B comparison testing.

🚀 Quick Start

# Activate the virtual environment
source .venv/bin/activate

# Check agent status
mab status

# Run a task with all agents in parallel
mab run "Write a Python validator function" --mode parallel

# List available workflows
mab workflows

# Compare agent configurations on a workflow
mab compare skill_development --skill "my_skill" --desc "Skill description"

🤖 Supported AI CLIs

All CLIs are pre-authenticated and ready to use:

Agent	Command	Path	Strengths
Codex	codex exec <prompt>	/usr/local/bin/codex	Code generation, refactoring
Claude	claude -p <prompt>	~/.local/bin/claude	Complex reasoning, architecture
Copilot	copilot -p <prompt>	/usr/local/bin/copilot	Git workflows, quick suggestions
Gemini	gemini -p <prompt>	/usr/local/bin/gemini	Research, multi-modal

📁 Project Structure

multi-agent-building/
├── orchestrator/           # Core Python package
│   ├── adapters.py         # Async CLI adapters (tested configs)
│   ├── cli.py              # Typer CLI (mab command)
│   ├── core.py             # Orchestrator with routing logic
│   ├── models.py           # Pydantic models
│   ├── workflows.py        # Workflow definitions
│   └── runner.py           # Workflow executor & comparison
├── scripts/
│   ├── discover_clis.py    # CLI discovery & validation
│   └── test_cli_configs.py # Agent configuration testing
├── skills/
│   ├── templates/          # Antigravity SKILL.md templates
│   └── developed/          # Generated skills output
├── config/
│   ├── agents.yaml         # Agent capabilities & routing
│   ├── cli_configs.json    # Tested CLI configurations
│   └── cli_discovery.json  # Discovered CLI paths
├── results/                # Workflow comparison results
└── pyproject.toml

🔄 Workflows

skill_development

Develop a new skill with research → design → implement → test → review.

Configurations:

• claude_heavy — Claude for complex tasks, Gemini for research
• codex_heavy — Codex for code tasks, Claude for design/review
• distributed — Each agent does what it's best at
• copilot_heavy — Copilot-centric workflow

code_review

Multi-agent code review: security → quality → refactoring suggestions.

Configurations:

• security_focused — Claude for security analysis
• balanced — Distributed review responsibilities

🔧 CLI Commands

# Agent management
mab status              # Show all agents and their status

# Task execution
mab run <prompt>        # Execute a task
  --type <type>         # Task type (code_generation, architecture, etc.)
  --mode <mode>         # single, parallel, sequential, consensus

# Skill development
mab develop <name>      # Start developing a new skill
mab skills              # List developed skills

# Workflow comparison
mab workflows           # List available workflows
mab compare <workflow>  # Run workflow with multiple agent configs
  --skill <name>        # Skill name for context
  --desc <description>  # Skill description
  -c <config>           # Specific configs to test (repeatable)

📋 Skill Output Format

Skills are generated in Antigravity-compatible format:

---
name: skill-name
description: What the skill does and when to use it
---

# Skill Title

Instructions in Markdown...

## When to Use
- Trigger conditions

## Instructions
Step-by-step guidance

## Examples
Code examples

🧪 Collaboration Modes

Mode	Description
single	One agent handles the entire task
parallel	Multiple agents work simultaneously, outputs merged
sequential	Chain of agents, each building on previous output
consensus	Agents vote/agree on best approach

⚙️ Tested CLI Configurations

Determined via scripts/test_cli_configs.py:

# Codex: Non-interactive code execution
codex exec "<prompt>" --skip-git-repo-check

# Claude: Print mode for scripting  
claude -p "<prompt>" --output-format text

# Copilot: Silent mode with auto-approval
copilot -p "<prompt>" --allow-all-tools -s

# Gemini: Yolo mode for auto-approve
gemini -p "<prompt>" -y --output-format text

📊 Comparison Results

Workflow comparisons are saved to results/ as JSON with:

• Per-step timing and agent used
• Success/failure status
• Output previews
• Summary metrics (fastest, most reliable)

License

MIT