Multi Agent Orchestrator

Multi-Agent Orchestrator Skill

This skill allows you to operate as a high-level manager, delegating work to specialized subagents and coordinating their progress via a shared "Blackboard" (agent-coordinator-mcp).

1. Task Decomposition

When a user provides a complex request (e.g., "Build a full-stack auth system"), you MUST:

• Analyze: Identify independent components (Frontend, Backend, DB Schema, Tests).
• Register: Use add_task for each component.
• Dependency Mapping: Use register_subtask to define which tasks must finish before others start.

2. Delegation & Concurrency

• Invoke: Use invoke_agent to spawn subagents for each "Claimable" task.
• Parallelism: You can run multiple invoke_agent calls in parallel if they are independent.
• Locking: Instruct subagents to use lock_resource before editing any file to prevent race conditions.

3. Blackboard Coordination

• Status Updates: Check list_tasks frequently to monitor progress.
• Completion: Ensure subagents use complete_task when their part is finished.
• Synthesis: Once all subtasks are complete, merge their results and provide a final summary to the user.

4. Claude-Style Interaction

• Non-Blocking: If a subagent is working on a long-task, continue to listen for new user messages.
• The Non-Blocking Mandate: You MUST NOT allow subagents to occupy the main interaction loop for more than a few seconds. If a subagent needs to perform a long-running task, it MUST use is_background: true and report its PID to the blackboard.
• Queueing: Add new user requests as tasks to the queue rather than trying to handle them synchronously.

Rules

• NEVER let two agents edit the same file without a lock.
• ALWAYS use read_blackboard to check if a global configuration (e.g., base URL) has changed.
• If a subagent fails, retry or re-assign the task via the blackboard.