Crew Manager Pattern

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.