Orchestrate Tasks: Advanced Multi-Agent Implementation

The orchestrate-tasks command provides advanced orchestration for complex features. Perfect for large features where you want fine-grain control over task group assignments, specialized agents, and targeted standards injection.

Choose implement-tasks OR orchestrate-tasks for a given spec—you wouldn't use both for the same feature.

How it works

Orchestrate-tasks creates an orchestration.yml file in your spec folder that serves as a roadmap for implementation. You'll specify:

• Task groups - Automatically detected from your tasks.md
• Agent assignments (if using Claude Code with subagents) - Which specialized subagent handles each task group
• Standards (if not using Claude Code Skills) - Which standards guide each task group's implementation

With Claude Code subagents

When use_claude_code_subagents is enabled, orchestrate-tasks delegates each task group to your specified Claude Code subagents. This provides efficient, parallel execution with isolated contexts.

You can assign any Claude Code subagents you've created. Learn how to create custom Claude Code subagents.

Tip: When creating custom Claude Code subagents, inject references to relevant agent-os/standards files in their instructions for better alignment with your coding standards.

Without Claude Code subagents

When not using subagents, orchestrate-tasks generates targeted prompts for each task group with fine-grain control over which standards are injected. This provides more precise context management compared to implement-tasks.

What gets created

• orchestration.yml - Your orchestration roadmap with task group assignments
• implementation/prompts/[task-group].md (if not using subagents) - Targeted prompts for each task group

How to run

Claude Code

Run command, /orchestrate-tasks

When using Agent OS with Claude Code commands enabled, run:

/orchestrate-tasks

The command will:

1. Create orchestration.yml with your task groups
2. Ask you to assign subagents to each task group (if using subagents)
3. Ask you to specify standards for each task group (if not using Claude Code Skills)
4. Delegate to subagents or generate prompts based on your configuration

Other Tools

Two ways to run this command

When using other AI coding tools, you have two options for running this command:

Option A: Run all steps in one shot

This single command instructs your agent to complete the entire orchestration process automatically. Some agents perform better at this than others.

@agent-os/commands/orchestrate-tasks/orchestrate-tasks.md run this

Option B: Run each numbered step manually

For more control and reliability, you can run each numbered command sequentially. Check your agent-os/commands/orchestrate-tasks/ folder for the numbered command files.

Either way, this command will generate numbered implementation prompts in agent-os/specs/[this-spec]/implementation/prompts/ that you run sequentially to implement each task group.

Example orchestration.yml

With Claude Code subagents and manual standards assignment:

task_groups:
  - name: authentication-system
    claude_code_subagent: backend-specialist
    standards:
      - all
  - name: user-dashboard
    claude_code_subagent: frontend-specialist
    standards:
      - global/*
      - frontend/components.md
      - frontend/css.md
  - name: api-endpoints
    claude_code_subagent: backend-specialist
    standards:
      - backend/*
      - global/error-handling.md

Next step

Congratulations! Your feature is now complete and ready for your final review and polish before shipping.

Important: Building with AI agents—with or without Agent OS—is rarely a one-and-done process. You should always review the implementation and add your final polish before shipping.

Your review should also serve as your opportunity to spot patterns in how your agents perform, so that you can optimize your standards and workflows for better performance and alignment.

Return to the workflow overview