Task decomposition: breaking work down so AI can act on it
Task decomposition turns a goal that is too big into a list of simple subtasks, which a model or agent then handles in order or in parallel. It is the planning step that lets an agent carry out long pieces of work without losing the thread.
Data as of October 22, 2026
Key figures
| Level | Advanced |
|---|---|
| Origin | 2022 |
| Introduced by | Zhou et al. (least-to-most prompting) |
What the term covers
Ask an AI agent to “write a comparison of three invoicing tools” and it cannot reliably produce the whole thing in one shot. Task decomposition makes it draw up a list first: identify the three tools, collect their prices, list their features, check compliance points, write the summary. Each subtask is small enough to be done properly, and its output feeds the next one.
There are two broad approaches:
- Fixed decomposition: the developer writes the steps in advance. This is the idea behind an agentic workflow: predictable and easy to test.
- Dynamic decomposition: the model builds the plan itself and revises it as it learns more. More flexible, but harder to keep under control.
Origins and key papers
Splitting a problem into sub-problems is an old idea in computer science and automated planning. For language models, a frequently cited milestone is “Least-to-Most Prompting Enables Complex Reasoning in Large Language Models” (Zhou et al., Google, May 2022, ICLR 2023). The model first breaks the question into easier sub-questions, then solves them in order, reusing each answer. It extends chain of thought by explicitly separating “plan” from “solve”.
Why it matters
Decomposition sits at the heart of agent orchestration: a lead agent splits the task and hands each piece to a sub-agent or a tool, as in a multi-agent system. It has three practical effects:
- Reliability: a small step is easier to get right, and to check, than a big one.
- Focused context: each subtask only needs the information relevant to it.
- Cost: more steps mean more model calls. The agent cost calculator helps you estimate what a ten-step plan costs compared with a two-step one.
Common myths
- “The more detailed the plan, the better.” Over-splitting multiplies calls and creates more hand-off points where things can break.
- “A plan made at the start is enough.” The best agents re-plan when a step fails or turns up something unexpected.
- “It’s only for agents.” Asking a chatbot to “list the steps first, then work through them one at a time” already applies the principle.
Related terms
- Chain of thought · Getting a model to write out intermediate reasoning steps before its final answer, which helps on multi-step problems.
- Agentic workflow · A sequence of steps in which an AI model plans, uses tools, checks and revises its work instead of answering in a single pass.
- Agent orchestration · Coordinating several model calls, tools or AI agents: who does what, in which order, and how the results are brought together.
- Multi-agent system · A setup where several AI agents, each with its own role, split a task between them and share their results to get it done.
- ReAct pattern · A loop in which an AI agent alternates written reasoning with actions such as searches or tool calls, reading each result before going on.
Try it
Full glossary
Frequently asked questions
Who does the decomposing, the model or the developer?
Either. In a workflow, the developer hard-codes the steps in advance. In an agent, the model proposes the plan, sometimes with a human approving it before anything runs.
Should every task be broken down?
No. For a simple question, splitting it up adds calls, time and cost without improving the answer. Decomposition pays off when a task has several dependent steps or independent parts that can run in parallel.
What happens if the plan is wrong?
The agent diligently carries out the wrong steps. That is why robust systems let the plan be revised along the way, based on intermediate results, instead of following it blindly.
Data sources
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.