Paper deep dive
OrgAgent: Organize Your Multi-Agent System like a Company
Yiru Wang, Xinyue Shen, Yaohui Han, Michael Backes, Pin-Yu Chen, Tsung-Yi Ho
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 4/2/2026, 3:32:26 AM
Summary
OrgAgent is a hierarchical multi-agent framework that organizes LLM-based agents into a company-style structure consisting of governance, execution, and compliance layers. By separating planning, task solving, and final verification, OrgAgent improves reasoning performance and token efficiency compared to flat multi-agent systems, particularly on complex benchmarks like SQuAD 2.0 and MuSiQue.
Entities (6)
Relation Signals (4)
OrgAgent → comprises → Compliance Layer
confidence 100% · a compliance layer for final answer control
OrgAgent → comprises → Governance Layer
confidence 100% · OrgAgent decomposes multi-agent reasoning into three layers: a governance layer...
OrgAgent → comprises → Execution Layer
confidence 100% · an execution layer for task solving and review
OrgAgent → improvesperformanceon → SQuAD 2.0
confidence 95% · the hierarchical setting improves performance over flat multi-agent system by 102.73%... on SQuAD 2.0
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:While large language model-based multi-agent systems have shown strong potential for complex reasoning, how to effectively organize multiple agents remains an open question. In this paper, we introduce OrgAgent, a company-style hierarchical multi-agent framework that separates collaboration into governance, execution, and compliance layers. OrgAgent decomposes multi-agent reasoning into three layers: a governance layer for planning and resource allocation, an execution layer for task solving and review, and a compliance layer for final answer control. By evaluating the framework across reasoning tasks, LLMs, execution modes, and execution policies, we find that multi-agent systems organized in a company-style hierarchy generally outperform other organizational structures. Besides, hierarchical coordination also reduces token consumption relative to flat collaboration in most settings. For example, for GPT-OSS-120B, the hierarchical setting improves performance over flat multi-agent system by 102.73% while reducing token usage by 74.52% on SQuAD 2.0. Further analysis shows that hierarchy helps most when tasks benefit from stable skill assignment, controlled information flow, and layered verification. Overall, our findings highlight organizational structure as an important factor in multi-agent reasoning, shaping not only effectiveness and cost, but also coordination behavior.
Tags
Links
- Source: https://arxiv.org/abs/2604.01020v1
- Canonical: https://arxiv.org/abs/2604.01020v1
Trouble viewing inline? Open PDF directly →
Full Text
51,547 characters extracted from source content.
Expand or collapse full text
OrgAgent: Organize Your Multi-Agent System like a Company Yiru Wang 1 Xinyue Shen 3 Yaohui Han 1 Michael Backes 3 Pin-Yu Chen 2 Tsung-Yi Ho 1 1 The Chinese University of Hong Kong, 2 IBM Research 3 CISPA Helmholtz Center for Information Security Abstract While large language model-based multi-agent systems have shown strong potential for com- plex reasoning, how to effectively organize mul- tiple agents remains an open question. In this paper, we introduce OrgAgent, a company- style hierarchical multi-agent framework that separates collaboration into governance, execu- tion, and compliance layers. OrgAgent decom- poses multi-agent reasoning into three layers: a governance layer for planning and resource al- location, an execution layer for task solving and review, and a compliance layer for final answer control. By evaluating the framework across reasoning tasks, LLMs, execution modes, and execution policies, we find that multi-agent systems organized in a company-style hierar- chy generally outperform other organizational structures. Besides, hierarchical coordination also reduces token consumption relative to flat collaboration in most settings. For example, for GPT-OSS-120B, the hierarchical setting im- proves performance over flat multi-agent sys- tem by 102.73% while reducing token usage by 74.52% on SQuAD 2.0. Further analysis shows that hierarchy helps most when tasks benefit from stable skill assignment, controlled infor- mation flow, and layered verification. Overall, our findings highlight organizational structure as an important factor in multi-agent reasoning, shaping not only effectiveness and cost, but also coordination behavior. 1 Introduction Large language models (LLMs) have evolved from single-turn assistants into increasingly autonomous agents capable of planning, tool use, and col- laboration. These advances have driven the de- velopment of LLM-based Multi-Agent Systems (MAS), which are increasingly studied in complex settings such as problem solving, software engi- neering, and simulation (Guo et al., 2024; Li et al., 2024; He et al., 2025). Existing research has de- veloped along two directions. One line studies HierarchicalStructure- GPT-5mini Task risk is high, so I set budget mode to HIGH with maximum rounds set to 3. LayerA-Governance I approve FULL MAS with reasoning as Drafter, domain as Specialist, and up to 3 rounds. LayerB-Execution Question: Who is the most likely murderer, Hugo or Maria? Context: Philip was found murdered with barbed wire in a library, and the case centers on the clues and alibis of two main suspects, Hugo and Maria. User I recommend reasoning as the Drafter, domain as the Specialist, and setting usespecialist = true. LayerC — Compliance CSO Hugo CCO PASS Format Check Drafter reasoning Selects Maria based on access and timing. Argues Hugo is more likely due to the barbed wire. Rejects answer and requests revision. Reviewer First Round Specialist domain Drafter reasoning Changes the answer to Hugo. AgreePass Reviewer Specialist domain Final Round ... MAS Hugo MASResponse 1#F81D6 2F8950 356CAE6 499D182 5E6F6E 6FFCB4C 7 Figure 1: Illustration of our company-style hierarchi- cal MAS framework OrgAgent. Layer A performs governance-level planning, including skill assignment and execution control; Layer B carries out task solving through collaborative drafting and feedback; Layer C finalizes the output through answer consolidation and compliance checking. interaction mechanisms among agents, focusing on how agents communicate and coordinate through role-playing, discussion, debate, voting, or consen- sus, as exemplified by CAMEL (Li et al., 2023). The other line focuses on higher-level organiza- tion, emphasizing role assignment, workflow de- sign, and system-level coordination, as represented by frameworks such as AutoGen (Wu et al., 2024) and role-specialized collaborative systems includ- ing MetaGPT (Hong et al., 2023), ChatDev (Qian et al., 2024), and Paperclip (paperclipai, 2026). One natural way to organize MAS is through organizational structure (Pugh, 1971; Mintzberg, 1979; Daft, 2007). In organizational theory, or- arXiv:2604.01020v1 [cs.MA] 1 Apr 2026 ganizational structure determines how tasks, coor- dination, supervision, and decision authority are distributed, thereby shaping organizational behav- ior (Burton et al., 2012). Common forms include flat structures (Ghiselli and Siegel, 1972) with fewer managerial layers and hierarchical struc- tures (Child, 2019) with more complex manage- ment. Among these, the company-style hierarchy has been refined over decades, developing well- established mechanisms for goal alignment, role division, resource allocation, and outcome verifica- tion (Mintzberg, 1979; Burton et al., 2012). This makes company-style hierarchy a natural basis for organizing MAS, as it explicitly defines who plans, who executes, who reviews, and how decisions are controlled. In this work, as shown in Figure 1, we instan- tiate organizational structure as a company style hierarchy, one of its most common real-world rep- resentations, to study how structured governance affects multi-agent reasoning. OrgAgent decom- poses the reasoning process into three layers: 1) a governance layer for planning, routing, and re- source allocation; 2) an execution layer for answer generation, critique, and revision, whose interac- tion process is further controlled through different execution modes and execution policies; 3) and a compliance layer for final answer validation and output control. We then evaluate the framework on three reasoning benchmarks, MuSR, MuSiQue, and SQuAD 2.0, using three language models and mul- tiple execution modes and policies. Results show that MAS organized in the company-style hierarchy generally outperforms both flat and single-agent baselines, with especially clear gains on MuSiQue and SQuAD 2.0. Our main contributions are as follows: • We propose OrgAgent, a company-style hier- archical MAS framework that separates gover- nance, execution, and compliance, supported by a skill-based worker pool and various exe- cution modes and policies. • We present the first systematic empirical study of flat and hierarchical MAS on general rea- soning tasks, treating organizational structure itself as the central variable of analysis. •We show that company-style hierarchy im- proves both effectiveness and efficiency in most settings, often achieving higher task per- formance while reducing token cost, with gains of up to +102.73% in F1-score and 74.52% fewer tokens on SQuAD 2.0. •We provide a fine-grained analysis of coordi- nation behavior, showing when hierarchical organization is effective and when it may be limited by additional overhead or coordination constraints. 2 Related Work Organizational Structure.Organizational struc- ture is a core concept in organization theory and organizational design (Joseph and Sengul, 2025; Mintzberg, 1979; Burton et al., 2012; Daft, 2007). Classic and contemporary work studies how differ- ent structural forms distribute authority, coordina- tion, and specialization, including functional, hier- archical, and matrix arrangements (Duncan, 1979; Galbraith, 1971; Kates and Galbraith, 2010; Anand and Daft, 2007). Recent reviews further identify configuration, control, channelization, and coordi- nation as major perspectives in organization design research (Joseph and Sengul, 2025). Empirical studies also show that structure matters in prac- tice, for example, in healthcare quality (Hearld et al., 2008) and manufacturing innovation and op- erational performance (Iranmanesh et al., 2021). However, this literature primarily concerns human organizations and does not explain how structure should be instantiated for within-task governance in LLM-based multi-agent systems. LLM-Based MAS. Prior work on LLM-based MAS mainly falls into two lines (Guo et al., 2024). One line studies local interaction and decision mechanisms. CAMEL uses role playing for au- tonomous cooperation (Li et al., 2023). Multi-agent Debate uses iterative critique to improve reason- ing (Du et al., 2023). Voting or Consensus shows that decision protocol, agent count, and discussion rounds can substantially affect performance (Kaes- berg et al., 2025). AgentVerse studies collaborative group composition and emergent behaviors (Chen et al., 2023). The other line studies higher-level or- chestration. AutoGen provides a general infrastruc- ture for multi-agent conversation (Wu et al., 2024). MetaGPT organizes agents through role special- ization and standardized procedures (Hong et al., 2023). ChatDev structures software development through specialized agents and chat chains (Qian et al., 2024). From Bits to Boardrooms proposes a hierarchical framework linking operational analysis Execution Policies STRICT / BALANCE / NOCAP / AUTO Input Question & Context Execution Mode DIRECT | LIGHT MAS | FULL MAS User Execution Policies •AUTO •STRICT •BALANCE •NOCAP Input Question & Context Format Check Strategic Direction& Coordination Execution Mode DIRECT | LIGHT MAS | FULL MAS LayerA–Governance Skill-Based Worker Pool •Technical •Quantitative •Reasoning •Domain •Communications •Data Resource Control •Task Risk •Maximum Rounds Technical RoleAssignment •Select Drafter •Decide Specialist Need •Select Specialist Select From Worker Pool LayerB-Execution Drafter Reviewer Specialist Drafter Reviewer Drafter DIRECT LIGHT MAS FULL MAS LayerC — Compliance CSOCCO Benchmark-Aligned Final Answer Format Check MASResponse MAS OrgAgent-HierarchicalStructure OrgAgent-HierarchicalStructure MAS Figure 2: Overview of OrgAgent, a company-style hierarchical MAS framework. with strategic decision making (Wang and Zhang, 2025). However, it remains unclear how organi- zational structure itself should govern within task coordination in MAS, which our work addresses. 3 OrgAgent In this section, we introduce OrgAgent, as shown in Figure 2, a multi-agent framework that orga- nizes collaboration through three layers of gover- nance, execution, and compliance. We describe its architecture, agent roles, and execution policies, which together define how coordination is struc- tured within a single task. 3.1Organizational Structures in Management Organizational structure refers to how authority, roles, and coordination are arranged within an orga- nization. In management research, different struc- tures shape how decisions are made and how work is controlled (Mintzberg, 1979). In this paper, we focus on two common forms in management: flat organization and a hierarchical organization, be- cause they represent two contrasting ways of orga- nizing collective work. Flat Organization. A flat organization refers to a flatter structure with reduced vertical differentia- tion. It is usually characterized by shorter commu- nication paths, less layered supervision, and greater autonomy. Its main advantage is flexibility (Re- itzig, 2022; Lee, 2022), while its main limitation is weaker control and less clear coordination in complex tasks. Hierarchical Organization.A hierarchical orga- nization refers to a structure with multiple levels of authority and clearer reporting relationships. It is characterized by stronger supervision, clearer role differentiation, and more structured coordina- tion. Its main advantage is control and accountabil- ity (Halevy et al., 2011), while its main limitation is slower communication and lower flexibility. 3.2 Agents We define the agents used in our system. The same set of agents can be reused under different orga- nizational settings, while their interaction patterns may vary across frameworks. Chief Executive Officer (CEO).The CEO (Van- cil, 1987) focuses on overall strategic direction and high-level coordination. Its role is to keep the problem-solving process aligned with the overall objective of the task. ChiefTechnologyOfficer(CTO). The CTO (Medcof, 2007) focuses on technical sound- ness and solution design. Its role is to examine the technical direction of the solution and ensure that the problem-solving process remains technically appropriate. ChiefOperatingOfficer(COO). The COO (Bennett and Miles, 2020) focuses on operational resources and execution efficiency. Its role is to consider resource usage, execution constraints, and the overall efficiency of the process. Drafter. The Drafter is the primary writer in the problem-solving process. Its role is to produce the main candidate answer and revise it when neces- sary. Reviewer.The Reviewer focuses on answer qual- ity and error detection. Its role is to examine the current draft, identify potential weaknesses or in- consistencies, and determine whether revision is needed. Specialist. The Specialist focuses on targeted support for difficult or error-prone parts of the task. Its role is to provide additional expertise or refine- ment when the main draft requires further support. ChiefSolutionsOfficer(CSO). The CSO (Wikipedia contributors, n.d.)is re- sponsible for producing the final answer under benchmark-specific constraints.Its role is to ensure that the final response matches the required answer format and task requirements of the target benchmark. ChiefComplianceOfficer(CCO). The CCO (British Columbia Securities Commission, n.d.) is responsible for checking whether the final output satisfies predefined structural requirements. Its role is to verify compliance with the required schema or output format, but it does not perform task reasoning itself. 3.3 Skill-Based Worker Pool Our framework maintains a pool of six skill-based workers, which can serve as either the Drafter or the Specialist during execution. • Technical: Focuses on implementation de- tails, procedural constraints, and structured problem solving. •Quantitative: Focuses on numerical, sym- bolic, and stepwise reasoning. •Reasoning: Focuses on logical consistency, multi-step inference, and chain coherence; •Domain: Focuses on domain-specific inter- pretation and contextual understanding. •Communications: Focuses on clarity, con- cise final phrasing, and answer presentation. • Data: Focuses on evidence extraction, pattern recognition, and information organization. These skill profiles are not tied to fixed bench- mark types. Instead, they provide reusable capa- bility orientations that can be instantiated under different execution roles depending on task needs. 3.4 Flat Framework We implement a flat framework in which all partic- ipating agents operate at the same organizational level. Specifically, the CEO, CTO, COO, Drafter, Reviewer, Specialist, and CSO interact as peer agents without an explicit layered chain of com- mand, and all of them work on the basis of shared task information and shared interaction context. Although these agents have different functional responsibilities, coordination, problem solving, an- swer checking, and final response generation are carried out within a single-level collaborative pro- cess. The CCO is not treated as a deliberative peer, but is used only for final structural compli- ance checking. 3.5 Hierarchical Framework Our hierarchical framework OrgAgent organizes agents into three vertical layers, namely Layer A, Layer B, and Layer C. This design separates high-level coordination, task execution, and final output control into different stages, so that the problem-solving process follows a structured top- down workflow rather than a single-level interac- tion process. Layer A (Governance Layer). Layer A is re- sponsible for high-level coordination and planning. It includes the CEO, CTO, and COO, which re- spectively focus on strategic direction, technical direction, and operational resources. Based on the task input, this layer determines the execution con- figuration for the downstream process. Layer B (Execution Layer). Layer B is respon- sible for task solving under the configuration de- termined by Layer A. It includes the Drafter, Re- viewer, and, when needed, the Specialist. In this layer, the Drafter produces the candidate answer, the Reviewer checks its quality, and the Special- ist provides targeted support for difficult or error- prone parts of the task. Layer C (Compliance Layer). Layer C is re- sponsible for final output generation and structural verification. It includes the CSO, which produces the final answer under benchmark-specific con- straints, and the CCO, which checks whether the output satisfies the required structural format. In this way, the final response is both benchmark- aligned and structurally compliant. 3.6 Execution Modes Our framework supports three execution modes: DIRECT, LIGHT MAS, and FULL MAS. These modes differ in how the execution layer is orga- nized and therefore provide different trade-offs be- tween efficiency, verification strength, and coordi- nation cost. DIRECT. In DIRECT configuration, the execu- tion layer relies on the Drafter to produce the can- didate answer directly, without invoking additional review or specialist support. This mode minimizes execution overhead and is suitable for relatively simple tasks or resource-constrained settings. LIGHT MAS. The LIGHT MAS configuration activates the Drafter and the Reviewer. In this set- ting, the Drafter first produces a candidate answer, and the Reviewer then checks its quality and deter- mines whether revision is needed. Compared with DIRECT, this mode introduces an additional ver- ification step while keeping the coordination cost relatively low. FULL MAS.The FULL MAS configuration acti- vates the Drafter, Reviewer, and Specialist. In ad- dition to answer generation and review, this mode allows targeted expert support for difficult or error- prone parts of the task. As a result, it provides the strongest execution support, but also incurs the highest coordination and computation cost. 3.7 Execution Policies Our framework supports four execution policies, namely STRICT, BALANCE, NOCAP, and AUTO. These policies control how aggressively the framework constrains resource usage and col- laboration during execution. STRICT. The strict policy emphasizes conser- vative execution by imposing tighter resource and interaction constraints. BALANCE.The balance policy provides a mid- dle ground between efficiency and execution sup- port. NOCAP. The no-cap policy minimizes explicit execution constraints and allows more flexible re- source usage when needed. AUTO.The auto policy adaptively selects an ex- ecution configuration according to task characteris- tics. 4 Experimental Setup In this section, we describe the experimental setup used to evaluate OrgAgent. 4.1 Models We evaluate our framework with three backbone LLMs: GPT-OSS-120B (Agarwal et al., 2025), GPT-5 mini (OpenAI), and Llama 3.1 8B (Ol- lama). These models represent different levels of capability, allowing us to examine whether the im- pact of organizational structure is universal across different models. 4.2 Benchmarks We evaluate the framework on MuSR, MuSiQue, and SQuAD 2.0, which cover different forms of reasoning difficulty. Additional benchmark details are provided in Appendix A.1. MuSR (Sprague et al., 2023)is a benchmark for multistep soft reasoning over long narratives, and we report accuracy as the evaluation metric. MuSiQue (Trivedi et al., 2022) is a benchmark for compositional multi-hop question answering, and we report standard F1 scores. SQuAD 2.0 (Li and Zhang, 2018) is a reading comprehension benchmark containing both answer- able and unanswerable questions, and we report standard F1 scores. 4.3 Evaluation Metrics We evaluate each system from three perspectives: task performance, token efficiency, and coordi- nation behavior. Task Performance.For MuSR, we report Accu- racy. For MuSiQue and SQuAD 2.0, we report the standard F1-score. LetNdenote the total number of evaluation examples andKthe number of re- peated runs for each setting. Since the framework is stochastic, each setting is run multiple times, and we report the mean and standard deviation across runs. Detailed definitions of the benchmark- specific metrics are provided in Appendix A.2. Model BaselineFlatHierarchical (AUTO) ∆ Improvement (%) ∆ Token Reduction (%) ScoreAvg tokenScoreAvg tokenScoreAvg token MuSiQue (F1-score) GPT-5mini51.28± 4.22,77850.31± 2.5028,47968.98± 1.7011,408+37.11%59.94% GPT-OSS-120B37.98± 2.582,68748.40± 1.5525,20957.58± 1.9812,046+18.97%52.22% LLaMA-3.1-8B 11.52± 3.092,37014.55± 0.0951,42532.59± 14.6512,322+123.99%76.04% MuSR (Accuracy) GPT-5mini29.00± 1.411,60362.45± 5.8013,41964.83± 2.877,195+3.81%46.38% GPT-OSS-120B50.65± 2.941,32869.00± 1.5412,70059.50± 1.085,994-13.77%52.80% LLaMA-3.1-8B10.33± 2.491,06137.41± 1.0925,60034.00± 0.715,912-9.12%76.91% SQuAD 2.0 (F1-score) GPT-5mini31.34± 0.9545828.77± 3.0715,68363.43± 1.513,245+120.47%79.31% GPT-OSS-120B26.61± 1.6042531.12± 0.0313,02163.09± 1.523,318+102.73%74.52% LLaMA-3.1-8B24.92± 2.6224028.17± 2.9422,80644.78± 3.035,188+58.96%77.25% Table 1: Performance and token cost comparison across baseline, flat, and hierarchical organizations. Token Efficiency.To measure coordination cost, we report the average token usage per example: AvgToken = 1 N N X i=1 t i ,(1) whereiindexes an evaluation example, andt i is the total number of tokens consumed for example i, including all agent interactions and final answer generation. To compare hierarchical and flat coordination, we further compute relative score improvement and token reduction: Improvement(%) = S hier − S flat S flat × 100, (2) T okenReduction(%) = T flat − T hier T flat × 100, (3) whereS hier andS flat denote the final task perfor- mance scores of the hierarchical and flat settings, respectively, andT hier andT flat denote their aver- age token usage. Coordination Behavior. We analyze coordina- tion behavior through the distribution of selected skill types for the Drafter and Specialist roles. For the unanswerable subset of SQuAD 2.0, we also re- port the abstention rate, defined as the proportion of unanswerable examples for which the system outputs a normalized no-answer response. Detailed definitions are provided in Appendix A.2. 5 Results In this section, we conduct extensive experiments to evaluate the effectiveness and coordination be- havior of our organizationally structured multi- agent framework. We aim to address the following research questions: 1) In general reasoning tasks, can hierarchical organization outperform flat MAS and single-agent baselines? 2) How do different organizational structures, execution modes, and execution policies trade off task accuracy against token cost? 3) What coordination patterns emerge under different organizational settings? 5.1 Performance of Different Structures We present the quantitative comparison of perfor- mance and token cost across baseline, flat, and hi- erarchical organizations in Table 1. The results in- dicate that hierarchical organization generally out- performs both flat MAS and single-agent baselines. This advantage is especially clear on MuSiQue and SQuAD 2.0, where the hierarchical setting achieves the best performance for all three models. On MuSiQue, hierarchical organization improves F1-score by 37.11% for GPT-5mini and 123.99% for LLaMA-3.1-8B, while also bringing an 18.97% gain for GPT-OSS-120B. On SQuAD 2.0, the gains are even larger, reaching 120.47%, 102.73%, and 58.96% over flat MAS for GPT-5mini, GPT-OSS- 120B, and LLaMA-3.1-8B, respectively. The results on MuSR are more mixed. Hierar- chical organization slightly outperforms flat MAS for GPT-5mini, but remains below the flat setting for GPT-OSS-120B and LLaMA-3.1-8B. This sug- gests that hierarchical coordination is not uniformly dominant across all reasoning tasks. At the same time, these two cases also show that when hier- archical coordination fails to translate additional structure into better answer quality, it may still re- quire substantial coordination cost in tokens. More importantly, hierarchical organization uses substantially fewer tokens than flat MAS in every setting. Compared with the flat organization, token usage decreases consistently across all three bench- marks and all three models, with reductions rang- ing from 46.38% to 79.31%. This reduction is not Execution Policies 0 25 50 75 100 F1-Score (%) 51.28% GPT-5 mini Execution Policies 0 25 50 75 100 37.98% GPT-OSS-120B Execution Policies 0 25 50 75 100 11.52% Llama-3.1-8B Execution Policies 0 25 50 75 100 Accuracy (%) 29.00% Execution Policies 0 25 50 75 100 50.65% Execution Policies 0 25 50 75 100 10.33% Execution Policies 0 25 50 75 100 F1-Score (%) 31.34% Execution Policies 0 25 50 75 100 26.61% Execution Policies 0 25 50 75 100 24.92% MuSiQue MuSR SQuAD 2.0 FLATAUTOSTRICTBALANCENOCAPBASELINE Figure 3: Performance comparison of different execution policies across three benchmarks. Rows correspond to MuSiQue, MuSR, and SQuAD 2.0, while columns correspond to GPT-5 mini, GPT-OSS-120B, and Llama-3.1-8B. Bars denote the performance under FLAT, AUTO, STRICT, BALANCE, and NOCAP policies, and the red dashed line indicates the single-agent baseline. marginal, but large and systematic: the hierarchical framework never increases token cost relative to flat MAS, and instead consistently cuts interaction overhead by nearly half or more. Overall, these results show that introducing explicit layers of gov- ernance, execution, and compliance can improve coordination quality while substantially reducing the communication cost of multi-agent reasoning. 5.2 Accuracy and Token Cost Trade off We compare different execution policies within the hierarchical framework in terms of task per- formance in Figure 3 and average token consump- tion in Table 2. We first examine how performance varies across policies, as shown in Figure 3, dif- ferent execution policies lead to distinct perfor- mance patterns across benchmarks and models. On MuSiQue, AUTO, BALANCE, and NOCAP gen- erally achieve stronger F1-scores than STRICT, indicating that allowing more flexible coordination is beneficial for this benchmark. In particular, for GPT-5 mini and GPT-OSS-120B, the best results are obtained under AUTO or BALANCE, while for Llama-3.1-8B, the execution policies all remain clearly above the baseline and flat setting. On MuSR, the differences among policies are smaller for GPT-5 mini, but become more visible for GPT- OSS-120B and Llama-3.1-8B. In these cases, no single policy dominates across all models, sug- gesting that policy effectiveness is more model- dependent on this benchmark. On SQuAD 2.0, the Avg Token ModelAUTOSTRICTBALANCENOCAPFLAT MuSiQue GPT-5mini11,5453,79512,71121,76628,479 GPT-OSS-120B12,0463,63312,42814,53925,209 LLaMA-3.1-8B12,3223,28212,39937,30648,198 MuSR GPT-5mini7,1953,2757,1289,29413,419 GPT-OSS-120B5,9943,1286,9537,06312,700 LLaMA-3.1-8B5,9122,1767,69116,63925,600 SQuAD 2.0 GPT-5mini3,2451,5543,2153,51315,683 GPT-OSS-120B3,3181,5393,1563,57413,021 LLaMA-3.1-8B5,1881,1483,67411,38522,806 Table 2: Average token consumption under different execution policies across benchmarks and models. performance gap among execution policies is rela- tively small, and all of them remain substantially stronger than the baseline and flat setting. This sug- gests that for SQuAD 2.0, the main benefit comes from adopting structured hierarchical coordination itself, while the choice of policy mainly affects efficiency rather than final accuracy. We next examine token consumption across ex- ecution policies. Table 2 shows a clear and con- sistent pattern: STRICT is the most token-efficient policy across all benchmarks and all three mod- els. On MuSiQue, STRICT reduces average token usage to 3,795, 3,633, and 3,282 for GPT-5 mini, GPT-OSS-120B, and Llama-3.1-8B, respectively, far below other execution policies. The same pat- tern holds on MuSR, where STRICT again uses the fewest tokens for all three models. On SQuAD GPT-5miniGPT-OSS-120BLLaMA-3.1-8B SQuAD 2.0 Reasoning Domain Communications Quantitative Technical Data Outer Ring: Specialist Inner Ring: Drafter Figure 4: Skill distribution on SQuAD 2.0 across GPT-5mini, GPT-OSS-120B, and LLaMA-3.1-8B. The pie charts show the proportion of selected skill profiles under the hierarchical framework. 2.0, the token advantage of STRICT is even more pronounced, with only 1,554, 1,539, and 1,148 av- erage tokens, respectively. By contrast, NOCAP is usually the most expensive execution policy, es- pecially on MuSiQue and MuSR, where its token usage grows substantially. AUTO and BALANCE generally occupy the middle range, offering mod- erate cost compared with STRICT and NOCAP. These results reveal a clear performance-cost trade-off across execution policies. A detailed de- scription of the relationships among the execution policies, particularly their differences in coordina- tion strictness and budget constraints, is provided in Appendix A.4. In contrast, STRICT consistently provides the strongest efficiency advantage, often preserving competitive performance while using only a small fraction of the token budget. Over- all, these results show that execution policy serves as an effective control knob within the hierarchi- cal framework: stricter policies favor efficiency, whereas more flexible policies can improve perfor- mance when additional coordination cost is accept- able. 5.3 MAS Coordination Behavior To better understand how different organizational settings shape collective reasoning, we further ana- lyze coordination behavior through the lens of skill- selection distributions and answer behavior. We primarily report results on SQuAD 2.0 in the main text, and provide further analysis on MuSiQue and MuSR in Appendix B. As shown in Figure 4, hierarchical coordination produces clear but model-specific specialization patterns on SQuAD 2.0. GPT-5mini and LLaMA- 3.1-8B overwhelmingly assign the drafter to the domain specialist, with selection rates of 87.50% and 90.82%, respectively. In contrast, GPT-OSS- 120B more often assigns the drafter to the reasoning specialist, reaching 73.50%. Specialist selection also differs substantially across backbones. GPT- 5mini relies most heavily on the data specialist, which accounts for 72.28% of specialist assign- ments. GPT-OSS-120B concentrates mainly on rea- soning and data specialists, at 50.94% and 32.08%. By contrast, LLaMA-3.1-8B distributes specialist assignments much more broadly across reasoning, domain, communications, quantitative, and tech- nical skills, indicating weaker specialization and less stable coordination. These results suggest that hierarchy provides a structured coordination mech- anism, but the resulting division of labor remains strongly shaped by the underlying model. In Table 3, the abstention results on the unan- swerable questions of SQuAD 2.0 further show how hierarchy changes system behavior in mean- ingful ways. Specifically, the single-agent base- line never abstains, and flat MAS shows little to no abstention, ranging from 0 to 3.02%. In con- trast, hierarchical execution policies raise absten- tion rates substantially, reaching 19.39% to 39.78%. Among these execution policies, NOCAP yields the highest abstention rates for all three models, reaching 31.18% for GPT-5mini, 39.78% for GPT- SQuAD 2.0 Unanswerable Questions Abstention Rate (%) GPT- 5mini GPT- OSS-120B LLaMA- 3.1-8B Baseline000 Flat3.0200 AUTO22.5836.5619.39 STRICT13.9832.2619.35 BALANCE30.1135.4821.51 NOCAP31.1839.7827.96 Table 3: AbsRate(%) on Unanswerable Questions in SQuAD 2.0 under different organizational settings and execution policies. OSS-120B, and 27.96% for LLaMA-3.1-8B. This pattern suggests that hierarchical coordination is especially useful when the task benefits from con- trolled information flow, stable role assignment, and layered checking, particularly in cases where the correct behavior is to withhold an answer rather than guess. 6 Conclusion We presented OrgAgent, a company-style hier- archical MAS framework that organizes collabo- ration through explicit layers of governance, exe- cution, and compliance. Across three reasoning benchmarks, we show that structuring agents like a company can often achieve a better balance be- tween task effectiveness and token efficiency than flat coordination and single-agent baselines, with especially clear gains on MuSiQue and SQuAD 2.0. We further find that hierarchy also changes coordination behavior itself, often leading to more structured role allocation and verification patterns. Our findings suggest that company-style hierarchy provides a useful paradigm for building more ca- pable, economical, and interpretable multi-agent systems. Limitations First, although OrgAgent performs well on open- ended reasoning tasks, its improvements are more limited on multiple-choice benchmarks such as MMLU (Hendrycks et al., 2020) and MMLU- Pro (Yue et al., 2025). A possible reason is that these tasks provide a constrained answer space, leaving less room for hierarchical coordination to contribute. Second, our framework uses a fixed maximum number of discussion rounds. When agents fail to converge before this limit, the in- teraction is forcibly terminated, which may leave the coordination process incomplete. In such cases, hallucinated or weakly supported claims introduced by one agent may not be fully corrected, and can instead be propagated or reinforced through subse- quent interaction. Although our hierarchical design includes review and compliance steps, it cannot fully eliminate this risk. Finally, we evaluate only a limited set of models, tasks, and organizational settings, and do not examine other practical factors such as latency, stability across repeated runs, or human evaluation. References Sandhini Agarwal, Lama Ahmad, Jason Ai, Sam Alt- man, Andy Applebaum, Edwin Arbus, Rahul K Arora, Yu Bai, Bowen Baker, Haiming Bao, and 1 others. 2025. gpt-oss-120b & gpt-oss-20b model card. arXiv preprint arXiv:2508.10925. Narasimhan Anand and Richard L Daft. 2007. What is the right organization design? Organizational dynamics, 36(4):329–344. Nate Bennett and Stephen Miles. 2020. Riding Shotgun: The Role of the COO, Updated Edition. Stanford University Press. BritishColumbiaSecuritiesCommission.n.d. Theroleofthechiefcomplianceoffi- cer.https://w.bcsc.bc.ca/industry/ registrant-regulation/compliance-toolkit/ role-of-the-chief-compliance-officer. Accessed: 2026-03-16. Richard M Burton, Gerardine Desanctis, and BøRge Obel. 2012. Organisational design: A step by step approach. Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chi-Min Chan, Heyang Yu, Yaxi Lu, Yi-Hsin Hung, Chen Qian, and 1 others. 2023. Agent- verse: Facilitating multi-agent collaboration and ex- ploring emergent behaviors. In The Twelfth Interna- tional Conference on Learning Representations. John Child. 2019. Hierarchy: A key idea for business and society. Routledge. Richard L Daft. 2007. Organization theory and design. Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. 2023.Improv- ing factuality and reasoning in language models through multiagent debate, 2023. URL https://arxiv. org/abs/2305.14325, 3. Robert Duncan. 1979. What is the right organization structure? decision tree analysis provides the answer. Organizational dynamics, 7(3):59–80. Jay R Galbraith. 1971. Matrix organization designs how to combine functional and project forms. Business horizons, 14(1):29–40. Edwin E Ghiselli and Jacob P Siegel. 1972. Leadership and managerial success in tall and flat organization structures. Personnel Psychology, 25(4). Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V Chawla, Olaf Wiest, and Xi- angliang Zhang. 2024. Large language model based multi-agents: A survey of progress and challenges. arXiv preprint arXiv:2402.01680. Nir Halevy, Eileen Y. Chou, and Adam D. Galinsky. 2011. A functional model of hierarchy: Why, how, and when vertical differentiation enhances group performance. Organizational psychology review, 1(1):32–52. Junda He, Christoph Treude, and David Lo. 2025. Llm- based multi-agent systems for software engineering: Literature review, vision, and the road ahead. ACM Transactions on Software Engineering and Method- ology, 34(5):1–30. Larry R Hearld, Jeffrey A Alexander, Irene Fraser, and H Joanna Jiang. 2008. How do hospital organiza- tional structure and processes affect quality of care? a critical review of research methods. Medical Care Research and Review, 65(3):259–299. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2020. Measuring massive multitask language under- standing. arXiv preprint arXiv:2009.03300. Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, and 1 others. 2023. Metagpt: Meta programming for a multi-agent collaborative framework. In The twelfth international conference on learning representations. Mohammad Iranmanesh, Kavigtha Mohan Kumar, Be- hzad Foroughi, Reza Kian Mavi, and Ng Hui Min. 2021. The impacts of organizational structure on op- erational performance through innovation capability: innovative culture as moderator: M. iranmanesh et al. Review of Managerial Science, 15(7):1885–1911. John Joseph and Metin Sengul. 2025. Organization de- sign: Current insights and future research directions. Journal of Management, 51(1):249–308. Lars Benedikt Kaesberg, Jonas Becker, Jan Philip Wahle, Terry Ruas, and Bela Gipp. 2025. Voting or consensus? decision-making in multi-agent debate. In Findings of the Association for Computational Linguistics: ACL 2025, pages 11640–11671. Amy Kates and Jay R Galbraith. 2010. Designing your organization: Using the STAR model to solve 5 criti- cal design challenges. John Wiley & Sons. Saerom Lee. 2022. The myth of the flat start-up: Re- considering the organizational structure of start-ups. Strategic Management Journal, 43(1):58–92. Guohao Li, Hasan Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. 2023. Camel: Communicative agents for" mind" exploration of large language model society. Advances in neural information processing systems, 36:51991–52008. Xinyi Li, Sai Wang, Siqi Zeng, Yu Wu, and Yi Yang. 2024.A survey on llm-based multi-agent sys- tems: workflow, infrastructure, and challenges. Vici- nagearth, 1(1):9. Yuanjun Li and Yuzhu Zhang. 2018. Question answer- ing on squad 2.0 dataset. In Department of Energy Resources Engineering and Department of Electrical Engineering. Stanford University. John W Medcof. 2007.Cto power. Research- Technology Management, 50(4):23–31. Henry Mintzberg. 1979. The structuring of organiza- tions. In Readings in strategic management, pages 322–352. Springer. Ollama. llama3.1.https://ollama.com/library/ llama3.1. Model library page. Accessed: 2026-03- 16. OpenAI. Gpt-5 mini model.https://developers. openai.com/api/docs/models/gpt-5-mini . OpenAI API documentation. Accessed: 2026-03-16. paperclipai. 2026. Paperclip: Open-source orchestra- tion for zero-human companies.https://github. com/paperclipai/paperclip. Accessed: 2026-03- 15. Derek Salman Pugh. 1971. Organization theory: Selected readings, volume 126.Penguin Har- mondsworth. Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng Su, Xin Cong, and 1 others. 2024. Chatdev: Com- municative agents for software development. In Pro- ceedings of the 62nd annual meeting of the associa- tion for computational linguistics (volume 1: Long papers), pages 15174–15186. Markus Reitzig. 2022. How to get better at flatter de- signs: considerations for shaping and leading organi- zations with less hierarchy. Journal of Organization Design, 11(1):5–10. Zayne Sprague, Xi Ye, Kaj Bostrom, Swarat Chaudhuri, and Greg Durrett. 2023. Musr: Testing the limits of chain-of-thought with multistep soft reasoning. arXiv preprint arXiv:2310.16049. Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2022.Musique: Multi- hop questions via single-hop question composition. Transactions of the Association for Computational Linguistics, 10:539–554. Richard F Vancil. 1987. Passing the baton: Managing the process of ceo succession. (No Title). Zihao Wang and Junming Zhang. 2025. From bits to boardrooms: A cutting-edge multi-agent llm framework for business excellence. arXiv preprint arXiv:2508.15447. Wikipedia contributors. n.d.Chief solutions offi- cer.https://en.wikipedia.org/wiki/Chief_ solutions_officer. Accessed: 2026-03-16. Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, and 1 others. 2024. Au- togen: Enabling next-gen llm applications via multi- agent conversations. In First conference on language modeling. Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, and 1 others. 2025. Mmmu- pro: A more robust multi-discipline multimodal un- derstanding benchmark. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15134– 15186. A Appendix A.1 Additional Details of the Benchmarks To complement the brief benchmark description in the main text, we provide additional details on the characteristics and scale of the three datasets used in our experiments. These benchmarks were se- lected because they stress different aspects of multi- agent coordination, ranging from long-context nar- rative reasoning to compositional evidence aggre- gation and answerability detection. Table 4 summa- rizes the three benchmarks used in our experiments, including their task types and overall dataset sizes. MuSR. MuSR is a benchmark for multistep soft reasoning over long free-text narratives. Rather than focusing on short factual lookups, it requires models to combine multiple pieces of implicitly distributed evidence and perform commonsense- driven reasoning over story-like contexts. The benchmark contains three domains: murder mys- teries, object placements, and team allocation. Ac- cording to the official benchmark description, these domains contain 250, 256, and 250 instances re- spectively, for a total of 756 examples. Because MuSR emphasizes long-form narrative understand- ing and non-trivial intermediate inference, it is es- pecially suitable for analyzing whether hierarchical coordination helps agents organize evidence and reduce reasoning errors in complex textual settings. MuSiQue. MuSiQue is a compositional multi- hop question answering benchmark designed to make shortcut-based reasoning difficult. Its con- struction explicitly combines single-hop questions into connected multi-hop questions, so the final answer depends on evidence drawn across multi- ple supporting paragraphs rather than on isolated lexical overlap. The official paper reports statis- tics for MuSiQue-Ans, the answerable version of the dataset: 19,938 training instances, 2,417 de- velopment instances, and 2,459 test instances, for a total of 24,814 examples. The same paper fur- ther notes that MuSiQue-Full contains twice as many questions in each split by pairing each an- swerable example with an unanswerable counter- part. In our setting, MuSiQue provides a useful testbed for studying whether organizational struc- ture improves multi-step evidence composition and answer synthesis under moderate context complex- ity. SQuAD 2.0.SQuAD 2.0 is a reading comprehen- sion benchmark that combines standard extractive QA with adversarially written unanswerable ques- tions. In contrast to purely answerable QA tasks, models must both extract a correct text span when one is supported by the passage and abstain when no answer is entailed. The official dataset statistics report 130,319 training examples, 11,873 devel- opment examples, and 8,862 test examples. The benchmark extends SQuAD 1.1 by adding over 50,000 unanswerable questions written to resemble answerable ones, making superficial span matching insufficient. This benchmark is particularly useful in our study because it tests whether structured co- ordination helps agents distinguish between answer generation and answer refusal, especially when plausible distractors are present in the context. A.2 Detailed Benchmark Metric Definitions Notation.LetNdenote the total number of eval- uation examples, and leti∈1, . . . , Nindex an example. For each examplei,ˆy i denotes the pre- dicted answer andy i denotes the corresponding gold answer. The indicator functionI(·)equals1if its condition is true and 0 otherwise. MuSR.For MuSR, we report Accuracy, defined as Accuracy = 1 N N X i=1 I(ˆy i = y i ).(4) This metric measures the proportion of examples for which the predicted answer exactly matches the gold answer. MuSiQue and SQuAD 2.0.For both MuSiQue and SQuAD 2.0, we report the standard token-level F1-score. LetP i andG i denote the predicted and gold answer token sets for examplei. We first compute precision and recall: Precision i = |P i ∩ G i | |P i | ,Recall i = |P i ∩ G i | |G i | , (5) where|P i ∩G i |is the number of overlapping tokens between the prediction and the gold answer,|P i |is the number of predicted tokens, and|G i |is the number of gold tokens. The example-level F1- score is F1 i = 2· Precision i · Recall i Precision i + Recall i ,(6) BenchmarkTypeTotal MuSRMultistep soft reasoning756 MuSiQueCompositional multi-hop QA24,814 SQuAD 2.0 Reading comprehension with unanswerable questions151,054 Table 4: Overview of the benchmarks. and the final benchmark-level F1-score is F1 = 1 N N X i=1 F1 i .(7) Across-run statistics. For each setting, we run the systemKtimes. Lets k denote the benchmark score obtained in runk, wherek ∈ 1, . . . , K. We report the mean score and standard deviation: ̄s = 1 K K X k=1 s k ,(8) std(s) = v u u t 1 K− 1 K X k=1 (s k − ̄s) 2 .(9) Abstention rate on unanswerable questions. For the unanswerable subset of SQuAD 2.0, let Udenote the set of unanswerable questions, letˆa i denote the system output for examplei, and letN denote the set of normalized no-answer outputs. The abstention rate is defined as AbsRate unans (%) = 1 |U| X i∈U I(ˆa i ∈N )× 100. (10) This metric measures how often the system abstains from answering questions that do not have a valid answer in the context. A.3 Framework Configurations and Maximum Rounds Table 5 summarizes the organizational settings used in our experiments. We consider two coordination structures: a flat organization and a hierarchical or- ganization. The flat setting does not impose explicit layered governance, and all agents interact within a single-level coordination process with a maximum of three rounds. By contrast, the hierarchical set- ting decomposes collaboration into distinct layers with different responsibilities. In Layer A, manage- ment agents (CEO, CTO, and COO) are responsible for high-level planning, role assignment, and execu- tion control, with up to three rounds of governance- level coordination. In Layer B, execution is carried out under three alternative modes with different StructureLayerConfigurationMax Round Flat–3 Hierarchical Layer ACEO / CTO / COO3 Layer B DIRECT1 LIGHT MAS3 FULL MAS5 Table 5: Organizational structures and maximum coor- dination rounds used in our framework. coordination depths: DIRECT, which produces an answer in a single round; LIGHT MAS, which allows lightweight iterative collaboration for up to three rounds; and FULL MAS, which supports deeper multi-agent interaction for up to five rounds. This design enables us to systematically vary both organizational structure and coordination depth, and to analyze how these choices affect task perfor- mance and token efficiency. A.4Relationships Between Execution Policies and Token Consumption This appendix provides a supplementary view of how execution policies relate to token consumption and performance. As shown in Figure 5, Figure 6, and Figure 7, STRICT is consistently the most token-efficient execution policy, while NOCAP usually uses the most tokens. AUTO and BAL- ANCE typically lie between these two extremes. The figures also show that the performance gain from additional tokens is benchmark dependent. On MuSiQue, more flexible policies often achieve stronger results, while on MuSR and SQuAD 2.0, the performance differences among execution poli- cies are smaller than their token differences. Over- all, these results suggest that the execution policies form a spectrum from efficiency-oriented coordina- tion to more flexible but more expensive coordina- tion. B Additional Coordination Pattern Analysis MuSiQue. As shown in Figure 8, skill selection on MuSiQue also exhibits clear model-dependent patterns. For GPT-5mini and GPT-OSS-120B, the drafter is dominated by the reasoning specialist, 05k10k15k20k50k Avg Token 0 25 50 75 100 F1-Score (%) // GPT-5 mini 05k10k15k20k50k Avg Token 0 25 50 75 100 F1-Score (%) // GPT-OSS-120B 05k10k15k20k50k Avg Token 0 25 50 75 100 F1-Score (%) // Llama-3.1-8B MuSiQue baseflatautostrictbalancenocap Figure 5: Token-performance trade-off on MuSiQue across GPT-5 mini, GPT-OSS-120B, and Llama-3.1-8B. 05k10k15k20k30k Avg Token 0 25 50 75 100 Accuracy (%) // GPT-5 mini 05k10k15k20k30k Avg Token 0 25 50 75 100 Accuracy (%) // GPT-OSS-120B 05k10k15k20k30k Avg Token 0 25 50 75 100 Accuracy (%) // Llama-3.1-8B MUSR baseflatautostrictbalancenocap Figure 6: Token-performance trade-off on MuSR across GPT-5 mini, GPT-OSS-120B, and Llama-3.1-8B. while specialist selection is concentrated on a small subset of skills, especially domain knowledge and, in some cases, quantitative or data support. In contrast, LLaMA-3.1-8B shows a more mixed al- location for both drafter and specialist skills, with responsibility distributed across several skill types. This again suggests that hierarchical coordination induces specialization, but the sharpness and sta- bility of this specialization depend strongly on the backbone model. MuSR.Figure 9 shows a similar trend on MuSR. GPT-5mini and GPT-OSS-120B continue to assign the drafter primarily to reasoning-oriented agents, while specialist usage is concentrated mainly on domain and data-related support. LLaMA-3.1-8B remains comparatively more diffuse, with special- ist assignments spread across multiple skills rather than concentrated on a single dominant type. Over- all, the MuSR results are consistent with the main text: hierarchy provides a structured mechanism for division of labor, but the resulting coordination pattern remains strongly model-specific. 05k10k15k20k Avg Token 0 25 50 75 100 F1-Score (%) GPT-5 mini 05k10k15k20k Avg Token 0 25 50 75 100 F1-Score (%) GPT-OSS-120B 05k10k15k20k Avg Token 0 25 50 75 100 F1-Score (%) Llama-3.1-8B SQuAD 2.0 baseflatautostrictbalancenocap Figure 7: Token-performance trade-off on SQuAD 2.0 across GPT-5 mini, GPT-OSS-120B, and Llama-3.1-8B. GPT-5miniGPT-OSS-120BLLaMA-3.1-8B MuSiQue Reasoning Domain Communications Quantitative Technical Data Outer Ring: Specialist Inner Ring: Drafter Figure 8: Skill-selection distributions across six skill types on MuSiQue. The top row shows the Drafter, and the bottom row shows the Specialist, across GPT-5mini, GPT-OSS-120B, and LLaMA-3.1-8B. GPT-5miniGPT-OSS-120BLLaMA-3.1-8B MuSR Reasoning Domain Communications Quantitative Technical Data Outer Ring: Specialist Inner Ring: Drafter Figure 9: Skill-selection distributions across six skill types on MuSR. The top row shows the Drafter, and the bottom row shows the Specialist, across GPT-5mini, GPT-OSS-120B, and LLaMA-3.1-8B.