L’interface est localisée ; les présentations et documents des Skills sont traduits automatiquement par IA dans votre langue. Le code, les commandes, chemins et termes techniques restent en version originale, et le document de l’auteur fait foi.
Explorer les Skills
Productivité · SKILL.md

crew-manager-pattern

crew-manager-pattern

Use the manager-worker pattern: a single manager agent decomposes, delegates, monitors, and aggregates while worker agents execute leaves — applied via your existing planning, assignment, and dispatch tools.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentation de la Skill

GitHub

name: crew-manager-pattern description: "Use the manager-worker pattern: a single manager agent decomposes, delegates, monitors, and aggregates while worker agents execute leaves — applied via your existing planning, assignment, and dispatch tools."

Crew manager pattern

Adapted from CrewAI's manager pattern (provenance:crewai). Applies inside PolyGents' existing planning + agent-directory + dispatch tooling — no separate orchestration framework needed.

When to use

  • A planning task with 3+ independent leaves that each fit a known agent template (front-end, back-end, research, ops).
  • A user request that arrives at the orchestrator and needs to be broken up and reassigned.
  • A sprint where the manager role (SCRUM Master, Shadow Engineering Lead) is doing the distribution.

The pattern

  • Manager (you). Owns the user conversation, the planning task tree, and the final synthesis. Never personally edits the same artefacts the workers are editing.
  • Workers. Each owns a single leaf or a tight set of leaves. They follow agent-status-pings, raise blockers via agent-handoff-protocol back to the manager, and run agent-self-check before declaring leaf done.

The loop

  1. Decompose. task-decomposition-tree. Each leaf gets an owner_template field naming which agent template should pick it up.
  2. Assign. agent_directory_list → planning_tasks_bulk_assign with the chosen digital_employee_id per leaf.
  3. Brief. For each worker, post an agent-handoff-protocol packet on the task description.
  4. Monitor. Watch the planning overview + the workers' ping cadence. If a worker goes dark for > 1 cadence, ping or reassign.
  5. Aggregate. When all leaves are done, the manager synthesizes the final user-facing answer and runs agent-self-check over the whole.
  6. Report. One reply to the user, not N.

Anti-patterns

  • Manager also does worker leaves. Pick one role per turn.
  • Workers reporting directly to the user without going through the manager.
  • Decomposition that the manager never re-checks for drift.

Tools

  • agent_directory_list, planning_tasks_bulk_assign, planning_overview, dispatch_task, agent-handoff-protocol, agent-status-pings.
Skills enregistrées