Task Decomposition Tree

Task decomposition tree

Most agent failure modes start with bad decomposition — either too coarse (a single mega-task that can't be reviewed) or too fine (a tree of trivial leaves that creates more coordination cost than work).

The structure

A decomposition tree has three levels:

• Root. The user's outcome, in their words. Don't editorialize.
• Branches. 3–7 mid-level deliverables that together produce the root.
• Leaves. Concrete, verifiable actions (1–4 hours of agent time each).

Each leaf has:

• An owner (you, another agent template, or a human).
• A definition of done (a sentence that says exactly when it's complete).
• An independent / dependent flag for parallel-agent-coordination.
• A short list of tools or skills required.

When to apply

• Any user request that requires more than 3 tool calls to satisfy.
• Any sprint or planning iteration.
• Any retrospective where you are figuring out why work missed.

The loop

1. Restate the root in one sentence.
2. Draft 3–7 branches. If you have more than 7, you are mixing levels; collapse some.
3. For each branch, draft leaves at the granularity where one agent (or one engineer) can finish it in a single sitting.
4. For each leaf, mark dependencies on other leaves (use leaf ids).
5. Persist via planning_tasks_bulk_create so the tree lives in the planning system, not just in the chat.

Anti-patterns

• One mega-leaf that hides 3 days of work behind a single status field.
• Leaves with no verifiable definition of done.
• Mixing levels — a "deploy to staging" leaf next to a "decide entire pricing model" leaf.