| name | multi-agent-workflow |
| description | Use when the user wants to plan, run, review, or explain a reusable coordinated agent workflow with Plan Packet Review, file ownership, isolated workspaces, worker task cards, integration, and review gates. |
Multi-Agent Workflow
Use this skill to coordinate agentic work without binding the process to one repository. Start with a short dispatch decision, then load only the reference needed for the task.
Startup
- Run
git status --short --branch.
- Check for a project profile or ask for one only if repository inspection cannot provide the missing constraints.
- Decide the task level:
L0, L1, L2, L3, or High-risk.
- Decide dispatch mode: single agent, light coordinated, standard coordinated, or full coordinated.
Required Output Before Dispatch
Task level:
Dispatch mode:
Use multiple agents:
Enabled roles:
Reason full workflow is not enabled:
Human review gates:
Allowed automatic continuation:
References
- Read
references/controller-runtime.md for the short operating rules.
- Read
references/workflow.md for role details, states, ownership, integration, review, and cleanup.
- Read
references/templates.md when creating prompts, task cards, or run records.
- Read
references/project-profile-template.md when a repository needs a project profile.
Guardrails
- Do not create worker threads or new workspaces before Plan Packet Review when the task is L2, L3, High-risk, or an explicit workflow dry run.
- Do not assign two workers to write the same file.
- Give workers only their role template, task card, relevant plan excerpts, and task entry files.
- Return to the user for scope expansion, high risk changes, failed required validation, publishing, or destructive cleanup.
- Keep small tasks small.