subgoal decomposition

Subgoal decomposition breaks a hard, sprawling task into smaller subgoals that can each be solved more directly, then composed back into the whole. A complex request like book me a trip to Kyoto is hopeless to attack in one leap, but splitting it into find flights, choose dates, reserve a hotel, build an itinerary makes each piece tractable. It is the divide-and-conquer instinct applied to open-ended agent tasks.

In an agent the decomposition is usually hierarchical: a high-level goal is expanded into ordered or partially-ordered subgoals, each of which may be solved directly, recursively decomposed further, or delegated to a specialized sub-agent. This mirrors hierarchical planning — the manager reasons about subgoals while workers handle primitive actions — and it shrinks the effective horizon the model must reason over at any one moment, which improves reliability and makes progress observable subgoal by subgoal.

The quality of the whole hinges on the decomposition itself: subgoals that are not truly independent, that omit a needed step, or that cannot be recombined will propagate failure no matter how well each piece is solved. Choosing the right granularity is a judgment call — too coarse and a subgoal is still too hard, too fine and coordination overhead and error accumulation dominate.