Hierarchical Decomposition
A planner recursively spawns sub-planners and workers: open-ended task breakdown where scope is unknown upfront.
A hierarchical-decomposition agent breaks an open-ended goal into sub-goals, spawning sub-planners that break those down further, recursively, until the leaves are small enough for workers to execute. Unlike a flat hub-spoke, the tree deepens as the problem demands, so it handles tasks whose scope is not known up front. It suits large, open-ended objectives where the decomposition itself is part of the work. In anatomy terms it is a planner core that recursively spawns planner and worker cores in a tree.
A top planner core decomposes the goal and spawns sub-planner cores, which recurse until leaf worker cores execute; results flow back up the tree; the perimeter gates each level's writes.
- ·Large, open-ended objectives where the scope is unknown up front.
- ·Complex research or build tasks that decompose into nested sub-tasks.
- ·Planning problems where the breakdown is the hard part.
- ·Handles open-ended scope by deepening the tree as needed.
- ·Each level stays simple: decompose or execute.
- ·A natural fit for genuinely hierarchical problems.
- ·Parallelism across sibling sub-trees.
- ·Deep trees compound latency and cost.
- ·A bad high-level decomposition derails everything below.
- ·Hard to bound: the tree can grow unexpectedly.
- ·Aggregating results up the tree adds complexity.
- Internal research (Scalarity), Agents vs Operators / Agent Archetypes v2 · unvalidated
Description, use cases and trade-offs: Field Book analysis, synthesised from established agent-architecture literature.