HTN-Orchestrate: A Hierarchical Task Network Framework for Multi-Agent Orchestration

Yanchen Deng; Jianguo Huang; Yao Long Teng; Yixuan Li; Handuo Zhang; Ning Wang; Liangcai Su; Xinyu Wang

cs.AIOct 15, 2026

HTN-Orchestrate: A Hierarchical Task Network Framework for Multi-Agent Orchestration

Authors: Yanchen Deng, Jianguo Huang, Yao Long Teng, Yixuan Li, Handuo Zhang, Ning Wang, Liangcai Su, Xinyu Wang,

Organizations: Nanyang Technological University, Singapore · Apodex.

Abstract

Orchestrating multiple agents on a complex task requires reconciling local subtask execution with global task planning. Dynamic dispatch approaches provide execution flexibility but do not explicitly model subtask hierarchies or dependencies, while predefined workflows impose this structure but fix the decomposition before execution begins. Therefore, we introduce HTN-Orchestrate, a two-layer framework that separates these responsibilities through a hierarchical task network (HTN). The Planner maintains subtask hierarchies and dependencies, deferring decomposition until prerequisite results become available, while local Orchestrators schedule workers to execute each leaf subtask. Results propagate upward under AND semantics, and accumulated findings inform a rebuilt task network when the current decomposition cannot resolve the task. On five benchmarks spanning web research, scientific reasoning, and research workflows, HTN-Orchestrate outperforms ReAct, Agent Team, and DeepSeek Harness using the same base model, achieving 89.73% accuracy on BrowseComp and 73.44% on SUPERChem, exceeding Agent Team by 2.45 and 4.69 percentage points, respectively.

Figures & tables

Figure 1: HTN-Orchestrate separates result-informed task planning from local multi-agent orchestration. (a) The Planner expands ready tasks using prerequisite results and synthesizes parent results after all children return. When rebuilding is needed, supported evidence is retained to construct a new task network. (b) Each executable leaf is delegated to an Orchestrator, which dispatches workers, collects reports as they become available, and uses those reports to guide subsequent assignments and leaf synthesis. The leaf result returns to the task network for further planning and upward synthesis.

	STR	RES	DEC	RPL	EVC
Agent Team ( Apodex Team, 2026 )		$\checkmark$
Workforce ( Hu et al., 2025a )	$\checkmark$		$\checkmark$	$\checkmark$
ADaPT ( Prasad et al., 2024 )	$\checkmark$			$\checkmark$
FlowSearch ( Hu et al., 2025b )	$\checkmark$	$\checkmark$		$\checkmark$
ChatHTN ( Muñoz-Avila et al., 2025 )	$\checkmark$
HTN-Orchestrate (ours)	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$

Table 1: Feature comparison with closely related systems. STR = structured plan (explicit task hierarchy or dependency graph); RES = result-informed (subsequent work conditioned on execution findings); DEC = decoupled (global planning and local execution handled by separate controllers); RPL = replanning (task structure revised during execution); EVC = evidence carryover (findings from a prior plan explicitly condition the rebuilt plan).

Proprietary
Method / Model	BC	FSR	FSO	SUPERChem (text-only)	PRL-Bench
GPT-6 Astra ( OpenAI, 2026c )	91.50	–	–	–	–
GPT-5.6 Sol ( OpenAI, 2026d )	90.40	–	–	73.60	–
GPT-5.5 ( OpenAI, 2026b )	84.40	33.90	73.80	66.20	–
GPT-5.4 ( OpenAI, 2026a )	82.70	23.33	–	57.70	37.38
Claude Fable 5 ( Anthropic, 2026a )	87.40	–	–	–	–

Table 2: Main results. BC, FSO, and SUPERChem report accuracy (%), FSR reports success rate (%), and PRL-Bench reports the mean percentage of satisfied rubric criteria. Higher is better for all metrics; bold marks the best score and underline marks the second best on each benchmark.

Panel of Figure 2 — Figure 2: Per-question completion times on a 200-question subset of BrowseComp, with marker shapes and colors indicating paired correctness. Both axes use logarithmic scales; points below the dashed equality line indicate shorter completion times for HTN.

Panel of Figure 3 — Figure 3: Worker concurrency on a 200-question subset of BrowseComp. (a) Average worker concurrency $\bar{C}$ over each question’s full trace window; the dashed line marks equality. (b) Per-question time shares, averaged equally over all 200 questions or the 165 questions both methods answer correctly. Workers are active from invocation to completion or cancellation, including model and tool waits; control agents are excluded.

Metric	Correct	Incorrect
	( $n=184$ )	( $n=16$ )
Final tree maximum depth	1.23	1.19
Final tree leaves	3.95	4.00
Executed leaf instances	4.52	6.31
Worker invocations	21.04	47.50
Orchestrator decision rounds	20.21	44.62

Table 3: HTN-Orchestrate’s internal statistics on a 200-question subset of BrowseComp, grouped by outcome. Tree metrics describe the final task network; execution counts include work before and after rebuilding. Values are means except for the percentage of questions with a rebuild.

Variant	Accuracy (%)	$\Delta$
HTN-Orchestrate	92.0	—
Upfront decomposition	90.5	$-1.5$
Single-agent leaf execution	89.5	$-2.5$
No rebuilding	89.5	$-2.5$

Table 4: Ablation results on a 200-question subset of BrowseComp.

\Delta

is the accuracy difference from HTN-Orchestrate in percentage points.

Panel of Figure 4 — Figure 4: Execution statistics on a 200-question subset of BrowseComp for HTN-Orchestrate (HTN) and its ablated variants. Panels (a–c) show per-question counts across all epochs. Boxes span the interquartile range, horizontal lines mark medians, and whiskers extend to 1.5 times the interquartile range; outlier markers are hidden. Panel (d) shows the percentage of questions with a task-network rebuild. Single-agent leaf execution is omitted from (c), and No rebuilding from (d), because their values are uniformly zero.

Panel of Figure 5 — Figure 5: Result-informed planning on BrowseComp Q193. The question traces an author of a 1993 role-playing-game supplement through a 2015 children’s book and his father’s radio program to a station’s first broadcast year. Solid lines show task hierarchy; dashed arrows show dependencies or returned evidence. Green boxes contain research findings. In (a), the blue subtree is expanded after the author list is returned. In (b), the orange tasks independently discover and reconcile author candidates within a tree fully planned before research begins. Both runs correctly identify WVIK and answer 1980.

Figure 6: Task-network rebuilding on BrowseComp Q501. The question asks for the third song in a 24-song blog list dated December 27, using a weekly-feature clue and the 21st song’s album as additional constraints. The orange node marks the initial incomplete post search; the blue node marks the added weekly-feature task. Green boxes show retained findings and subsequent results. Solid lines show hierarchy, and dashed arrows show selected dependencies and evidence transfers. Gray dashed nodes were planned but not executed; the final answer is supported by the completed post-search branch in epoch 1.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Benchmark	Questions	Task format	Retrieval
BrowseComp	1,266	Multi-hop web research; short factual answers	Enabled
BC ablation subset	200	Same task format; fixed question IDs across variants	Enabled
FrontierScience-Research	60	Research-level scientific problems; written solutions	Disabled
FrontierScience-Olympiad	100	Olympiad-level problems; worked solutions and final answers	Disabled
SUPERChem	256	Text-only chemistry questions; option selection	Disabled
PRL-Bench	100	Research-level physics problems; written solutions	Disabled

Appendix

Table 5: Evaluation sets and solver tools. Python execution is available in all five settings. Offline evaluation disables search, page retrieval, and external network access while retaining Python computation.

Setting	Value
Temperature / top- $p$	1.0 / 0.95
Maximum output tokens per model call	65,536
Maximum tree depth	2
Maximum children per decomposition	4
Maximum Orchestrator decisions per leaf per epoch	12
Worker capacity per leaf	3

Appendix

Table 6: HTN-Orchestrate execution settings. Tree depth counts the root as depth zero. Worker capacity applies within a leaf; independent ready leaves may execute concurrently.

Field	Meaning
id , parent_id , children	Node identity and hierarchical edges
title , task	Short label and self-contained objective
depends_on	Sibling nodes whose completion is required before this node becomes ready
needs_decomposition	Whether a ready node returns to the Planner
status	pending , running , completed , or failed
result_summary , ref	Evidence summary and reference to the full leaf result

Appendix

Table 7: Task-node representation.

Role	Input	Output
Planner	Root question, current task, completed precursor summaries, restart context	Direct execution or validated child tasks and dependencies
Orchestrator	Leaf objective, prerequisite evidence, returned reports, active assignments, budget	Dispatch, wait, or terminate
Instruction generator	Leaf objective, selected role, evidence and remaining issue	An executable worker assignment
Researcher	Assigned objective, original constraints, prior evidence	Findings, evidence, unresolved constraints, confidence
Verifier	Proposed answer, original objective, supporting reports	Completeness and correctness checks, verdict
Synthesizer	Node objective and worker or child reports	Evidence summary and completion outcome

Appendix

Table 8: Role inputs and outputs.

Variant	Changed behavior	Retained components
Upfront decomposition	Complete the task network before executing any leaf in each epoch. Decomposition instructions refer to anticipated outputs rather than completed precursor results.	Dependency-constrained execution, local multi-agent orchestration, synthesis, rebuilding
Single-agent leaf execution	Launch exactly one ReAct researcher for each executable leaf. Its report goes directly to leaf synthesis, replacing Orchestrator decisions and multi-worker dispatch.	Result-informed planning, prerequisite context, leaf and tree synthesis, rebuilding
No rebuilding	Set the maximum number of rebuilds to zero.	Result-informed planning, local multi-agent orchestration, synthesis, finalization from accumulated evidence

Appendix

Table 9: Components changed by each ablated variant.

Method	Correct	Acc. (%)	Mean time	Med. time	Exec. leaves	Workers	Decisions	Rebuilt
HTN-Orchestrate	184	92.0	58.03	25.23	4.665	23.155	22.165	35
Upfront decomposition	181	90.5	67.69	30.02	5.025	24.760	24.025	45
Single-agent leaf execution	179	89.5	28.49	17.69	6.145	6.145	0	64
No rebuilding	179	89.5	46.54	28.15	4.120	19.010	18.500	0

Appendix

Table 10: BrowseComp-200 ablation results. Execution counts and duration include all epochs of each selected attempt. Time is measured in minutes; leaves, workers, and decisions are means per question.

Ablated variant	Both correct	HTN correct only	Ablated variant correct only	Both incorrect	Net HTN gain
Upfront decomposition	176	8	5	11	$+3$
Single-agent leaf execution	172	12	7	9	$+5$
No rebuilding	175	9	4	12	$+5$

Appendix

Table 11: Paired correctness on BrowseComp-200. “HTN correct only” means the full system is correct and the ablated variant is incorrect; “Ablated variant correct only” denotes the reverse.

Figure 7: Exclusive correct outcomes in the ablation comparison. Blue bars count questions answered correctly only by HTN-Orchestrate; orange bars count questions answered correctly only by the ablated variant. The difference gives the net gain in Table 11 .

Figure 8: Local multi-agent orchestration on BrowseComp Q1025. The expanded leaf shows complementary discovery assignments followed by verification of a returned candidate. The full system identifies Linus Lerner; the single-agent-per-leaf variant submits Gregory Wait.

Question	Method	Leaves	Workers	Decisions	Rebuilds	Answer
Q193	HTN-Orchestrate	4	15	18	0	1980
Q193	Upfront decomposition	5	20	21	0	1980
Q1025	HTN-Orchestrate	3	10	10	0	Linus Lerner
Q1025	Single-agent leaf execution	9	9	0	1	Gregory Wait
Q501	HTN-Orchestrate	5	72	56	1	Serial Killer
Q501	No rebuilding	2	18	18	0	Silent Night

Appendix

Table 12: Execution counts for the three BrowseComp case pairs. Counts cover the entire selected question attempt.

Multi-Agent Orchestration

Multi-agent orchestration coordinates specialized AI agents, tools, and workflows to solve complex tasks efficiently and reliably.

See all 206 papers

Decoupled Multi-Agent Orchestration
Oct 6, 2026Xinle Wu, Yao LuMulti-Agent Orchestration LLM Agent Orchestration
National University of Singapore
Uno-Orchestra: Parsimonious Agent Routing via Selective Delegation
May 6, 2026Zhiqing Cui, Haotong Xie, Jiahao Yuan +11Multi-Agent System Optimization Multi-Agent Orchestration
Project lead.
AdaptOrch: Task-Adaptive Multi-Agent Orchestration in the Era of LLM Performance Convergence
Feb 18, 2026Geunbin YuMulti-Agent Orchestration LLM Agent Orchestration
Department of Artificial Intelligence, Korea National Open University

Oct 6, 2026J/K move · Enter open · S save

Decoupled Multi-Agent Orchestration

Xinle Wu, Yao Lu

National University of Singapore

Multi-Agent Orchestration LLM Agent Orchestration Multi-Agent Planning Task Decomposition

Open paper AskPDF arXiv

Learned orchestration can automatically construct effective language-model multi-agent systems, but existing approaches couple planning to fixed worker pools and train decomposition and collaboration from the same terminal outcome, limiting transfer and obscuring credit assignment. We introduce DeOrch, which separates worker-agnostic planning from concrete worker selection. Its two-stage planner first decomposes the task without worker information, then chooses collaboration operations using compact, worker-identity-free matchability feedback from the pool, enabling conditional credit assignment to decomposition and collaboration decisions. A lightweight matcher estimates worker suitability from behavior on a fixed probe set and adapts online with a contextual bandit, allowing new workers to be incorporated without retraining the planner or matcher. Across diverse in- and out-of-distribution tasks, DeOrch outperforms prior automatic MAS orchestration methods with fewer worker calls than competing learned orchestrators, remains effective when transferred to an entirely unseen worker pool without retraining, and shows consistent gains from both components.

Abstract

Multi-Agent Orchestration

Explore similar work

Decoupled Multi-Agent Orchestration

Uno-Orchestra: Parsimonious Agent Routing via Selective Delegation

AdaptOrch: Task-Adaptive Multi-Agent Orchestration in the Era of LLM Performance Convergence