Learn · Agentic architecture & orchestration · CCAR-F task 1.6

Task decomposition

Perfect subagents can't save a narrow plan. Cover the whole question, size the pieces, fix the output format, and choose fixed vs dynamic.

Key ideas

• Make subtasks cover the whole question; check coverage before delegating.
• A report with gaps usually means the coordinator's decomposition was too narrow, not that the subagents failed.
• Independent parts run in parallel; dependent parts run in order.
• Right-size the pieces and give each subtask a defined output format (findings, sources, confidence).
• Known, predictable steps: fixed prompt chain. Open-ended: dynamic decomposition.
• For big unknown jobs, map first, prioritise, then revise the plan as you learn.

Worked example: Narrow vs complete

For "Should we adopt AI for month-end close?", tools + costs is too narrow. Add accuracy, audit trail and controls; data security; peer adoption with measured results; and risks with a rollback plan. When a draft still shows a gap, the coordinator re-delegates that area.

Exam traps

• Blaming or re-prompting the subagents when whole areas are missing. The fix is the coordinator's decomposition.
• Using dynamic, agent-led decomposition for a review that runs the same three checks on every PR.
• Splitting work into tiny subtasks whose coordination cost outweighs the work.
• Letting each subagent return a different format, so results can't be merged.

Official docs

• Building effective agents
• Subagents in the Agent SDK

Practise CCAR-F questions · Next lesson: Sessions: resume and fork