Decomposition: Splitting the Work
The word decomposition sounds like something from chemistry class, where a compound breaks apart into simpler elements. In planning, decomposition means exactly the same thing: taking a complex goal and breaking it into simpler pieces until each piece is small enough to actually do. It is the single most powerful planning move that both humans and AI agents can make.
The Core Idea: Divide to Conquer
When you face a big goal, ask: what are the major parts of this task? Each major part becomes a subgoal — a smaller goal that contributes to the big one. Then you do the same thing to each subgoal: what are its parts? You keep dividing until you reach steps that are small enough to actually execute in one move. This process is called a decomposition tree, because the original goal sits at the top like a tree trunk, and the branches keep splitting downward into smaller and smaller tasks until the leaves are individual executable actions. For example: write a short film. Decompose into: write the script, plan the shoot, record the footage, edit the final video. Now decompose write the script into: brainstorm the story idea, outline the scenes, write each scene's dialogue, revise. Each of those can be decomposed further until you have a list of actual, doable tasks.
Decomposition is the process of breaking a complex goal into progressively smaller subgoals until each piece is small enough to execute directly. It transforms an overwhelming task into a manageable tree of work.
One of the earliest and most successful AI planning systems was called STRIPS, developed in 1971. STRIPS worked by decomposing a goal into a sequence of subgoals and checking which actions in its repertoire could satisfy each one. The robot it controlled could navigate rooms and move objects by solving this decomposition tree — not by doing anything magical, just by repeatedly answering the question: what smaller thing must be true before the bigger thing can be true?
Decomposition in Modern AI Agents
Today's AI agents use decomposition in a more flexible way. A large language model acting as an agent might receive a goal like: research and summarize the top five climate-change solutions proposed in the last two years. Instead of attempting to produce a full answer immediately, a well-designed agent first decomposes the task: identify what needs to be searched, perform each search, extract the relevant findings, compare the results, write the summary. This decomposition happens in a planning phase before any web search or writing begins. The agent builds an internal outline of work, then executes that outline step by step. If one search comes back empty, the agent can update the plan — not panic and start over. The best modern agent frameworks, like ReAct (Reason + Act) and chain-of-thought prompting, are essentially formal decomposition systems: they force the agent to articulate what it plans to do, one step at a time, before doing it.
The temptation — for humans and agents alike — is to start doing before planning is done. Resist it. A few seconds or tokens spent decomposing a goal saves enormous wasted effort later.
Match each term about decomposition to its correct meaning.
Terms
Definitions
Drag terms onto their definitions, or click a term then click a definition to match.
How Deep Should You Decompose?
A common question is: when do you stop splitting? The answer is: stop when you reach a step that is directly executable — meaning you or the agent can do it right now without further planning. If you write organize my research and you are not sure how to start, that is still too big. Split further. If you write open browser and search for solar panel efficiency 2024, that is directly executable — you know exactly what to do. That is a leaf. Over-decomposing is also a risk. Breaking brush teeth into lift toothbrush, apply paste, place on lower left molar, move up and down is absurdly granular — the overhead of managing hundreds of tiny steps outweighs the benefit. Good decomposition finds the level where steps are clear without being trivial.
Complete the description of a good decomposition.
What is the main benefit of decomposing a goal before acting on it?
A student wants to learn guitar. They decompose this into: practice scales, learn chords, play songs. They then decompose learn chords into: learn G, learn C, learn D, learn Em. At what point should they stop decomposing?
Build a Decomposition Tree
- Step 1: Choose a complex goal from this list, or invent your own: plan a fundraiser for a local cause, create a 5-minute documentary about your neighborhood, or design and launch a class podcast.
- Step 2: At the top of a blank page, write the main goal.
- Step 3: Draw two to four branches below it. Each branch is a major subgoal. Label them.
- Step 4: For each subgoal, draw two to three more branches — the smaller tasks inside it. Label those too.
- Step 5: Circle every task at the bottom (the leaves). For each leaf, write yes next to it if you could start doing it RIGHT NOW with no further planning, or no if it still needs splitting.
- Step 6: For any no items, add one more level of branches. When every leaf is a yes, your tree is complete.