Managed Agent Orchestration

Instructions

Use this skill when you are setting up or iterating on Claude Managed Agents that use memory, outcomes (rubric + grader), multiagent orchestration (lead + specialists), and/or webhooks.

1. Define the run goal and constraints

• Write a short success statement (1–3 sentences).
• List non-goals and constraints (latency, cost, safety, data access).

2. Write an outcomes rubric

• Use the rubric template in templates/outcomes-rubric.md.
• Keep criteria testable: each criterion should be answerable with Yes/No or a small ordinal scale.
• Add “failure modes” that the grader should look for.

3. Plan multiagent orchestration

• Identify whether you need a lead agent and which specialist agents are required.
• Keep specialists narrow and tool-focused; avoid overlapping scopes.
• Use the role glossary in references/role-glossary.md to keep terminology consistent.

4. Add a dreaming/memory review loop

• After runs, review what the agent learned and what it missed.
• Use references/dreaming-review-checklist.md to capture recurring mistakes, preferred workflows, and user-specific preferences.

5. Integrate webhooks (if needed)

• If your product needs async completion notifications, add a webhook for “outcome done” events.
• Treat webhook delivery as best-effort and design idempotent handlers.

Examples

Example: outcomes rubric skeleton

See templates/outcomes-rubric.md.

Example: role mapping

See references/role-glossary.md.

Example: dreaming review checklist

See references/dreaming-review-checklist.md.

Source

• https://claude.com/blog/new-in-claude-managed-agents