Cognition · Software component

HTN Planner

Software componentCognitionCognition & Memoryarc:HTNPlanner

A task planner that recursively replaces abstract tasks in a task network with subtask networks from applicable decomposition methods until only primitive, executable tasks with consistent constraints remain.

Responsibility. Decomposes a high-level goal through method-based hierarchical refinement into a primitive executable task network.

Also known as: Hierarchical Task Network planner, Hierarchical planner, Decomposition engine

Variant of Task Planner abstract

When to choose. Choose when the goal has naturally nested structure, reusable decomposition patterns exist, stakeholders need multiple abstraction views, or staged commitment under uncertainty is needed, and the world is stable during planning.

Direct neighbourhood (hover for relationship types)

Relationships

is configured by structural

Decomposition Method Library Ch5.4

delegates to dependency

invokes dependency

is invoked by dependency

Replanner abstract Ch5.4

reads dependency

alternative to variability

Design guidance

SHOULD be used when tasks are naturally nested, decomposition patterns recur, several stakeholders need different abstraction levels, or high-level decisions must precede detailed commitment.
SHOULD use 3-5 abstraction levels; deeper hierarchies add method-selection, propagation and state-projection overhead.
SHOULD NOT be used for simple 3-10 step sequential workflows, highly dynamic domains, optimization-dominant problems, or domains lacking decomposition knowledge.
SHOULD organise abstraction levels by decision relevance, not by time scale.
MAY be augmented with local optimization within each abstraction level when optimality matters.

Quantitative guidance

As stated by the sources; verify before use.

Planner reasons about 3-7 items per level instead of hundreds of primitive actions (Ch5.4 travel example).
5 abstract subtasks with 4 applicable methods each yields 4^5 = 1,024 search branches under systematic search (Ch5.4).
Four-level hierarchies may need 10x more planning time than two-level hierarchies for simple problems; 3-5 levels suit most domains (Ch5.4).
Final manufacturing primitive network may contain 50-100 concrete operations (Ch5.4).

Classification

Patterns: Hierarchical Task Network (HTN) planningDepth-first task selectionBreadth-first task selectionLeast-commitment task selectionGreedy method selectionSystematic backtracking searchInformed (A*) method searchStrategic/tactical/operational abstraction levelsLocal repair
Technologies: LangGraphAutoGenCrewAISemantic Kernel planner
Quality attributes: Performance efficiency (ISO/IEC 25010)Maintainability (ISO/IEC 25010)Explainability (NIST AI RMF: explainable and interpretable)Interaction capability (ISO/IEC 25010)
Risks mitigated: Combinatorial search-space explosion of flat planningPremature detailed commitmentOpaque unmaintainable flat plans

Sources

Ch5.4: T. Nguyen, "Hierarchical Planning Fundamentals," in Mastering Agentic AI Systems: Guide for the NVIDIA NCP-AAI Exam, 1st ed. 2026, ch. 5.4. ISBN: 9798244538229.