| name | multi-agent-orchestration |
| description | Activate this skill when delegating tasks, defining subagents, or orchestrating parallel agentic workflows. |
Multi-Agent Orchestration & Workflow Guidelines
To improve efficiency and prevent the primary developer agent from becoming a bottleneck, workflows should utilize specialized subagents for concurrent execution.
1. Concrete Subagent Profiles
When delegating tasks, use the define_subagent and invoke_subagent tools to create and manage the following profiles:
-
Codebase Researcher:
- Role: Dedicated to mapping out large existing codebases before major refactors.
- Tasks: Runs
grep_search and view_file to collect context, read documentation, and summarize architectural patterns.
-
Test Specialist:
- Role: Dedicated to writing unit test suites and enforcing coverage gates.
- Tasks: Generates test files, runs
Vitest coverage commands, and ensures the strict 45% coverage gate is met independently of feature development.
-
Documentation Auditor:
- Role: Dedicated to reviewing and maintaining project documentation.
- Tasks: Reviews markdown files, updates timestamps (using the
append_timestamps.py or check_docs_review.py scripts), and formats documentation.
2. Orchestration Constraints
The AI agent MUST adhere to the following Active Structural Constraints when managing multi-agent pipelines:
- Subagent Invocation: Before performing time-consuming, parallelizable tasks (e.g., broad codebase research, running a full test suite while writing code), you MUST explicitly execute
invoke_subagent to spawn the appropriate profile (Codebase Researcher, Test Specialist, or Documentation Auditor), assign them a clear objective, and END your turn. You are FORBIDDEN from performing these specialized tasks sequentially if they can be delegated.
- Asynchronous Synchronization: After invoking a subagent, you MUST NOT use loop-polling to wait for completion. You MUST proceed with other parallelizable work or END your turn to yield to the system until you receive an asynchronous message containing the subagent's result.
3. Sequential Workflows
When implementing sequential multi-agent pipelines (where agents operate one after another), the AI agent MUST adhere to the following Active Structural Constraints:
- Pipeline Setup: You MUST pass output artifacts (e.g.,
implementation_plan.md or research_notes.md) as input context to the next subagent in the sequence.
- Role Handoffs: When an agent completes its task, it MUST explicitly define how the next agent should resume work, ensuring a seamless handoff (e.g., "The Planner has finished the design; the Implementer should now execute the code modifications").
- Workspace Sharing: Sequential subagents MUST use
Workspace="share" or Workspace="inherit" during invocation to ensure they can access and modify the previous agent's file system changes.
Last Updated: 2026-07-22 | Last Reviewed: 2026-07-22