jaggerxtrm/orchestrating-agents

Orchestrating Agents

This skill provides a structured workflow for multi-model collaboration between Gemini and Qwen, enabling deep "handshaking" sessions for complex tasks.

Usage

Command-Line Invocation (Gemini)

# With workflow parameter (recommended)
/orchestrate collaborative "Design authentication system with JWT"
/orchestrate adversarial "Review payment processing security"
/orchestrate troubleshoot "Debug production memory leak"
/orchestrate handshake "Quick review of rate limiting logic"

# Without parameter (interactive selection)
/orchestrate "Design authentication system"

Workflow Types

| Workflow | Turns | Use Case | |----------|-------|----------| | collaborative | 3 | New features, architecture planning, design decisions | | adversarial | 3 | Security audits, vulnerability assessment, red-team review | | troubleshoot | 4 | Complex debugging, production errors, emergency fixes | | handshake | 1 | Quick second opinion, simple validation |

Interactive Selection (Fallback)

If no workflow type is specified, the agent will prompt for selection:

For Gemini CLI (ask_user)

ask_user({
  questions: [{
    question: "This task requires deep collaboration. Which orchestration workflow should we use?",
    header: "Orchestrate",
    options: [
      { label: "Collaborative Design", description: "Agent A proposes -> Agent B critiques -> Agent A refines." },
      { label: "Adversarial Review", description: "Agent A proposes -> Agent B attacks (Red Team) -> Agent A defends." },
      { label: "Troubleshoot Session", description: "Multi-agent hypothesis testing for emergency bugs." },
      { label: "Single Handshake", description: "Standard one-turn second opinion." }
    ]
  }]
});

For Claude Code (AskUserQuestion)

AskUserQuestion({
  questions: [{
    question: "This task requires deep collaboration. Which orchestration workflow should we use?",
    header: "Orchestrate",
    options: [
      { label: "Collaborative Design", description: "Agent A proposes -> Agent B critiques -> Agent A refines." },
      { label: "Adversarial Review", description: "Agent A proposes -> Agent B attacks (Red Team) -> Agent A defends." },
      { label: "Troubleshoot Session", description: "Multi-agent hypothesis testing for emergency bugs." },
      { label: "Single Handshake", description: "Standard one-turn second opinion." }
    ]
  }]
});

Workflows

1. Handshake Protocol (Standard)

| Agent | Initial Command | Follow-up (Resume) | | :--- | :--- | :--- | | Gemini | gemini -p "..." | gemini -r latest -p "..." | | Qwen | qwen "..." | qwen -c "..." |

2. Multi-Turn Loops

For complex scenarios, follow the turns defined in references/workflows.md.

Loop Progress:
- [ ] Turn 1: Proposal (Agent A)
- [ ] Turn 2: Cross-Examination (Agent B)
- [ ] Turn 3: Final Synthesis (Agent A)

Resources

• Workflows: See references/workflows.md for turn-by-turn logic.
• Config: Patterns and priority are defined in config.yaml.
• Examples: See references/examples.md for concrete patterns.
• Handover: See references/handover-protocol.md for command flags.

Best Practices

• Concise Context: Send only relevant snippets, not entire files, to the neighboring agent.
• Specific Tasks: Ask for a "Security Audit" or "Refactoring Suggestions" rather than a general review.
• Fail-Safe: If a neighboring agent is not found, proceed with a self-review but notify the user.