Org-Agent: Beyond Personal Assistants Towards Organizational Agents

Luyao Zhuang; Yujing Zhang; Zijin Hong; Yilin Xiao; Xiao Huang

cs.AISep 28, 2026

Org-Agent: Beyond Personal Assistants Towards Organizational Agents

Authors: Luyao Zhuang, Yujing Zhang, Zijin Hong, Yilin Xiao, Xiao Huang

Organizations: The Department of Computing, The Hong Kong Polytechnic University, Hong Kong SAR

Abstract

Language model agents serving organizations must coordinate requests from multiple users while using knowledge distributed across their interactions. We identify two complementary capabilities for this setting, namely cross-user interaction and decision-making, as well as cross-user memory and knowledge use. Both capabilities are governed by organizational constraints across three aspects: user identity, authority, and access permissions; the attribution and temporal validity of information; and rules for resolving conflicting requirements across users and completion requirements for joint decisions. These constraints shape what information or decisions must be obtained before an action can proceed and what conditions must be satisfied during its execution. Motivated by this, we introduce Org-Agent, a unified constraint-centric reasoning framework that organizes task execution in three stages. Specifically, Org-Agent decomposes a task into atomic subtasks and constructs a task dependency graph whose edges encode the dependencies among them. Building on this graph, it schedules the subtasks in dependency order through topological sorting. It then executes each subtask while accounting for the task's constraints, supported by evidence-acquisition and memory-management tools. Experiments on MUSES-Bench and GroupMemBench demonstrate the effectiveness of Org-Agent on both capabilities, and ablations further support the contributions of dependency modeling and tool use.

Figures & tables

Figure 1: Two complementary capabilities of agents for organizations. Left: cross-user interaction and decision-making , illustrated by coordinating a product launch based on multiple users’ requests. Right: cross-user memory and knowledge use , illustrated by answering a question about the login method using prior interaction history.

Figure 2: Overview of Org-Agent . Given user inputs, interaction history, and organizational constraints, ❶ TDG construction decomposes the task into subtasks and represents their dependencies in a Task Dependency Graph; ❷ dependency-aware scheduling orders the nodes through topological sorting; and ❸ constraint-aware node execution uses the LLM to determine and take actions toward node objectives based on the assembled execution context, with evidence-acquisition tools providing relevant records and memory-management tools supplying memory for context assembly and updating it with information from actions and observations.

RAG-based Methods
Method	Multi-Hop	Update	Ambiguity	Implicit	Temporal	Abstention	Avg.
BM25	36.26	15.89	14.15	42.86	41.36	68.35	37.72
text-embedding-3-large	32.42	12.15	23.58	34.69	25.31	66.91	33.29
HippoRAG2	28.57	18.69	14.15	34.69	33.33	79.86	36.11
GraphRAG	8.24	0.93	6.60	14.29	1.85	78.42	19.06
Agent Memory Systems

Table 1: Results (%) of baselines and Org-Agent on GroupMemBench across six query categories using GPT-4o-mini. The best result for each category is highlighted in bold , while the second best is indicated with an underline . Avg. denotes the average performance.

Model	Method	Multi-user Instruction		Cross-user Access		Meeting	Avg.
Model	Method	Queue ( $F_{1}$ )	Instruct (Acc.)	Privacy	Utility	Success Rate	Avg.
	Vanilla	62.97	58.69	96.94	64.96	37.04	64.12
	Org-Agent (Ours)	74.24	73.41	97.59	56.52	60.19	72.39
GPT-4o-mini	$\Delta\uparrow$	$+11.27$	$+14.72$	$+0.65$	$-8.44$	$+23.15$	$+8.27$
	Vanilla	78.16	81.18	71.26	97.35	75.93	80.78
	Org-Agent (Ours)	88.21	83.78	75.25	99.17	78.70	85.02

Table 2: Results (%) of Vanilla and Org-Agent on MUSES-Bench across three backbones. The best result for each metric within each backbone is shown in bold . Avg. denotes the average performance.

\Delta\uparrow

denotes the difference between Org-Agent and Vanilla in percentage points.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Benchmark	Category	# Samples
GroupMemBench	Multi-Hop	182
	Update	107
	Ambiguity	106
	Implicit	49
	Temporal	162
	Abstention	139

Appendix

Table 3: Statistics of the benchmarks in evaluation. Samples correspond to questions for GroupMemBench and scenarios for MUSES-Bench.

Component	Specification
GPU	$4\times$ NVIDIA GeForce RTX 4090 (24 GB each)
	$4\times$ NVIDIA GeForce RTX 4090 D (24 GB each)
CPU	$2\times$ Intel Xeon Gold 6426Y (16 cores each)

Appendix

Table 4: Hardware configuration of our local experimental server.

Benchmark / Task	Example node
GroupMemBench	Find User_13’s decision about the UAT validation scope
MUSES-Bench: Queue	Decide whether to accept Alice’s instruction to prioritize urgent requests
MUSES-Bench: Instruct	Respond to Alice in English using at most three sentences
MUSES-Bench: Cross-user Access	Handle Bob’s request for the project budget
MUSES-Bench: Meeting	Ask David whether he can attend on Friday at 14:00

Appendix

Table 5: Examples of task-specific nodes in Org-Agent across GroupMemBench and MUSES-Bench. Each node represents a subtask with a concrete objective.

RAG-based Methods
Method	Multi-Hop	Update	Ambiguity	Implicit	Temporal	Abstention	Avg.
BM25	36.81	28.04	18.87	36.73	38.89	41.73	34.36
text-embedding-3-large	29.12	22.43	22.64	42.86	17.28	33.09	26.31
HippoRAG2	29.67	23.36	17.92	38.78	24.07	64.03	32.89
GraphRAG	7.69	0.93	5.66	6.12	4.32	51.80	13.83
Agent Memory Systems

Appendix

Table 6: Results (%) of baselines and Org-Agent on GroupMemBench across six query categories using Qwen3-8B. The best result for each category is highlighted in bold , while the second best is indicated with an underline . Avg. denotes the average performance.

Explore similar work

Organizational Principles Enable Collective Intelligence in Embodied AI
Sep 10, 2026Zhengran Ji, Jonathan Hyun, Boyuan ChenOrganizational Research Embodied Artificial Intelligence
Department of Computer Science, Duke University, Durham, USA. · Department of Electrical and Computer Engineering, Duke University, Durham, USA. · Department of Mechanical Engineering and Materials Science, Duke University, Durham, USA.
Can LLM Agents Sustain Long-Horizon Organizational Dynamics?
May 31, 2026Xuancheng Zhu, Yang Yue, Shuaibing Wan +4Long-Horizon Agents Multi-Agent Simulations
Beijing University of Posts and Telecommunications, Beijing, China
Focus Is All You Need: Adaptive Goal-aware Attention Orchestration for Multi-Agent Graph Systems
Jul 26, 2026Mingzhou Fan, Siyuan Xu, Mingxuan YuanReasoning Paths

Sep 10, 2026, cs.MAJ/K move · Enter open · S save

Organizational Principles Enable Collective Intelligence in Embodied AI

Zhengran Ji, Jonathan Hyun, Boyuan Chen

Department of Computer Science, Duke University, Durham, USA. · Department of Electrical and Computer Engineering, Duke University, Durham, USA. · Department of Mechanical Engineering and Materials Science, Duke University, Durham, USA.

Organizational Research Embodied Artificial Intelligence Intelligence

Open paper AskPDF arXiv

Collective intelligence depends not only on the capabilities of individual members, but also on how those members are organized. Yet artificial multi-agent systems are typically assembled using fixed organizational structures, even when the physical tasks they perform impose fundamentally different coordination requirements. Here we show that principles from human organization theory can be operationalized to organize large, heterogeneous collectives of embodied artificial agents. We introduce ORCH (Organizing Roles and Coordination Hierarchies), which constructs task-specific hierarchical organizations by combining pooled interdependence for work that can proceed concurrently with sequential interdependence for work governed by prerequisite relationships. Across 25 wildfire-response missions spanning reconnaissance, rescue, transportation, resource management, containment and suppression, we evaluated teams of up to 50 heterogeneous agents using eight large language models. Organizations constructed using these principles consistently outperformed four representative embodied multi-agent approaches across mission outcome, execution efficiency, exploration and computational resource use. Human-designed ORCH organizations improved final score by 63.97% and execution efficiency by 74.29% on average relative to the four prior frameworks. Organizations generated automatically by language models improved these measures by 43.63% and 52.53%, respectively. These advantages persisted across missions and underlying language models. Notably, collective performance was not monotonically determined by model scale. Analysis of long-horizon missions showed that hierarchical organization enabled teams to preserve concurrent activity within specialized groups while coordinating ordered transitions between mission phases.

Abstract

Explore similar work

Organizational Principles Enable Collective Intelligence in Embodied AI

Can LLM Agents Sustain Long-Horizon Organizational Dynamics?

Focus Is All You Need: Adaptive Goal-aware Attention Orchestration for Multi-Agent Graph Systems