The outcome

Gain useful specialization without multiplying coordination errors, permissions, latency, token cost, and debugging complexity.

Step by step

A workflow you can repeat.

  1. 01

    Benchmark one agent on a frozen task set and identify specific failures in expertise, context, parallelism, routing, or deterministic control that justify a team or workflow.

  2. 02

    Give each member one role, minimal knowledge and tools, explicit handoff schema, separate budgets and permissions, and choose coordinate, route, broadcast, or workflow control deliberately.

  3. 03

    Keep consequential effects in deterministic approval-gated steps, cap delegation depth and calls, isolate member context, validate handoffs, and prevent members from expanding their own capabilities.

  4. 04

    Test wrong routing, contradictory members, delegation loops, prompt injection propagation, permission escalation, partial failure, stale shared state, cancellation, latency, and total cost.

  5. 05

    Compare quality and trajectory against the single-agent baseline, inspect member traces, remove members without measurable value, canary changes, and retain a simpler fallback.

Working standard

What good use looks like.

  • Prove the single agent is insufficient.
  • Give each member minimal permissions.
  • Measure coordination cost and failures.

Official references

Check the current product documentation.