Playbook

1. Draw the Task Graph

Mark dependencies and shared files. Parallelize only nodes that can produce independently verifiable outputs.

2. Choose the Topology

• Use coordinator-worker for bounded fan-out and centralized synthesis.
• Use peer communication only when workers must negotiate during execution.
• Use an external orchestrator for durable state, retries, scheduling, or cross-session workflows.

Choose from documented product semantics, not from the generic “team” label.

3. Allocate Ownership

Give each worker a unique task ID, file scope, input snapshot, output contract, and deadline. Define who may change shared files.

4. Bound the Run

Limit active workers, nesting depth, retries, tokenTokenA chunk of text — roughly three-quarters of a word. The unit the model reads and bills in.The fragments a model breaks text into. Around 100 tokens ≈ 75 English words. Limits and pricing are counted in tokens, not words or characters.Full definitionClose or credit budget, and wall-clock duration. Prefer one deliberate fan-out over recursive uncontrolled spawning.

5. Reconcile Centrally

One accountable coordinator reviews results, resolves conflicts, and reruns integrated verification. Peer agreement is not acceptance evidence.

6. Measure Against One Agent

Track elapsed time, total model usage, defects found after integration, merge effort, and recovery from failure. Keep the team only when it improves the system outcome.