Task Decomposition — AI Glossary

Reference details

Related terms

• DAG (Directed Acyclic Graph)
• Parallel Execution (Agents)
• Least-to-Most Prompting

How it works

1. The goal is read against the available tools, because a sub-task the agent has no way to perform is not a decomposition but a wish.
2. Sub-tasks are emitted with explicit dependencies, which is what distinguishes a graph from a list and determines what may run in parallel.
3. Each sub-task carries its own completion criterion, so the agent can tell done from attempted without re-reading the whole goal.
4. Execution follows a topological order of the graph; independent branches may be dispatched together.
5. The plan is revised when a step returns something the plan did not anticipate — a decomposition produced once and never revisited is the common failure, not a wrong first plan.

Example: Where the dependency edge earns its keep

A goal needs five sub-tasks: A and B independent at 4s each, C depends on A at 3s, D depends on B at 3s, E depends on C and D at 2s. Run strictly in sequence that is 4+4+3+3+2 = 16s. Respecting only the real edges, A and B run together (4s), then C and D together (3s), then E (2s): 9s. The saving comes from the edges that are absent, which is why a flat list of five sub-tasks costs 16 seconds no matter how good each one is.

Commonly confused with

One step within planning: turning a goal into a set of sub-tasks, often as a list or a dependency graph. It is not the same as the plan being good — a correct decomposition executed in the wrong order, or never revised, still fails. Where the output is a graph, it is a DAG, which is what makes independent branches parallelisable.

When to use it

Reach for it when:

• A goal needs several steps whose order is not fixed in advance by code.
• Some sub-tasks are independent, so a dependency graph permits parallel dispatch.
• Progress has to be reportable mid-run, which requires named sub-tasks with their own criteria.

Reach for something else when:

• The steps are always the same, in which case a fixed workflow is cheaper and testable.
• The task is a single call — decomposition adds a planning round-trip and nothing else.
• Nothing can verify a sub-task's completion, leaving the agent to grade its own progress.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Task Decomposition — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path