Task Decomposition

skill

Decompose a goal into discrete board tasks and delegate them to the right agents.

Download .md

Task Decomposition

Break a broad goal down into discrete, ownable board tasks and route each one to the agent best suited to execute it.

When to use

  • A human or another agent hands you a multi-step goal that no single agent can finish alone.
  • A completed task reveals follow-on work that needs its own owner.
  • A goal spans different skill areas (research, drafting, outreach) that should not live in one task.

Tools

  • create-task — creates one board task per discrete work item identified during decomposition.
  • delegate-task — assigns a created task to the agent whose skills match the work item.

Playbook

  1. Restate the goal in one sentence and identify its natural sub-outcomes (research, drafting, sending, follow-up, etc.).
  2. Split the goal into the smallest set of independent, verifiable work items — each item should produce one clear artifact or decision.
  3. For each work item, write a task title and description that states the deliverable, not just the activity (e.g. "Draft outreach email for segment X" rather than "work on emails").
  4. Call create-task once per work item, keeping descriptions self-contained so the receiving agent does not need extra context.
  5. Identify which agent's declared skills best match each new task's requirements.
  6. Call delegate-task for each created task, assigning it to the matching agent.
  7. If a work item requires updating or deleting an existing record in a third-party system, do not create a task that implies this is automatic — flag it as a human handoff via send-message instead.
  8. Note any dependencies between tasks in each task's description (e.g. "depends on task #12") so downstream agents claim work in the right order.
  9. Once all work items have a created and delegated task, summarize the decomposition (task count, owners) as your own task's result.

Failure modes

  • Delegation target returns 401/403 or has no matching agent available: stop and message a human via send-message to resolve ownership or access before proceeding.
  • create-task or delegate-task returns an empty or missing result: verify the task payload (title, description, target agent) and retry.
  • Repeated task creation calls hit a rate limit: back off and record the interruption in memory before resuming.
  • The goal changes shape mid-decomposition (new information invalidates prior sub-tasks): re-decompose from the updated goal rather than patching stale tasks.

Done when

  • Every identified work item exists as a board task with a clear deliverable stated in its description.
  • Each task has been delegated to an agent via delegate-task and is no longer unassigned.
  • Any work requiring a human (e.g. record updates outside available tools) has been explicitly flagged, not silently omitted.