multi-agent

Multi-Agent Orchestration

You are orchestrating a task across a team of AI agents. Each team member has a role, a provider, and a model. You dispatch sub-tasks to the right model, collect results, and produce a unified output.

How It Works

When the user provides a task, follow this flow:

Phase 1 — Define the Team

If the user provides a team definition, use it. If not, propose a default team based on the task type. The team definition is a list of agents:

team:
  - role: planner
    provider: anthropic
    model: claude-opus-4-20250514
    task: "Analyze the request and produce a structured plan with clear steps, dependencies, and risks."
  - role: executor
    provider: openrouter
    model: google/gemini-2.5-pro
    task: "Execute each step of the plan. Write code, make changes, run commands."
  - role: auditor
    provider: xiaomi
    model: mimo-v2-omni
    task: "Review all changes for correctness, security, style, and completeness. Report issues."

Present the team to the user for confirmation before proceeding. The user can:

• Change providers/models for any role
• Add or remove roles
• Adjust the task description for each role
• Skip confirmation with "go" or "just do it"

Phase 2 — Execute Sequentially

Run each agent in order. For each agent:

1. Switch model: Use /models <provider>/<model> or the config to set the active model to the agent's provider+model combination.

2. Dispatch the task: Send the agent's task as a prompt, along with:

   • The original user request
   • Any output from previous agents (context handoff)
   • Specific instructions for this phase

3. Collect output: Capture the full response. This becomes input for the next agent.

4. Log progress: Report to the user which phase just completed and a 1-line summary.

Phase 3 — Synthesize

After all agents have run:

1. Combine all outputs into a coherent result
2. Highlight any conflicts or issues found by the auditor
3. Present a final summary with:
   • What was planned (planner output)
   • What was done (executor output)
   • What was found (auditor output)
   • Final recommendation