Task Decomposition
Ask an agent to prepare the quarterly sales report and the model has no action available to it — the sentence names an outcome, not a move. Task decomposition is the step that closes that gap: break the request into pieces small enough that each one maps onto something the agent can actually do with the tools it has.
When the split happens
There are two timings, and the choice shapes the whole run:
- Up front — the agent writes the full breakdown before acting, as in Plan-and-Execute. Cheap, auditable, and inspectable before anything irreversible happens; brittle the moment reality differs from the plan.
- As it goes — each piece is chosen in light of the last observation, as in ReAct. Adapts to surprises, but drifts, and repeats itself once early steps have fallen out of the context window.
Production systems usually do both: a coarse plan held as explicit state, with each piece decomposed further on arrival. Decomposition is also how work is handed out — an orchestrator splits a job and dispatches pieces to worker agents, which only pays off when the pieces are genuinely independent.
Getting the grain right
The failure modes come as a pair. Too fine, and the agent spends its step budget on bookkeeping — four calls to read one file. Too coarse, and a piece like clean the data fails as a unit with no information about which part broke, so the retry repeats all of it. Good pieces have a single tool call at their core, a result the agent can check, and no hidden dependency on a step that has not run yet.
Task decomposition is the breaking of a high-level request into pieces small enough that each maps onto an action the agent can take with its tools. It happens up front, as in Plan-and-Execute, or step by step from the latest observation, as in ReAct, and it is also how work is distributed to worker agents. Granularity decides the run: too fine wastes the step budget, too coarse makes failures impossible to localise.
Nobody moves house in one action. You book a van, pack the kitchen, label the boxes, hand back the keys. A big job only becomes doable once it is cut into pieces you can start on.
Task decomposition is an AI agent cutting up its instructions the same way. Handed something broad — sort out the quarterly report — it works out the list of smaller jobs that add up to it, then takes them one at a time.
The skill is in the size of the pieces. Too small, and the agent fusses over trivia until it runs out of turns. Too big, and when a piece goes wrong it has no idea which part of it failed, so it starts the whole thing again.