Paper deep dive
Too Many Specialists: Emergent Inefficiencies and Bottlenecks for Multi-agent Ad-hoc Collaboration
Benjamin Panny, Shashank Mehrotra, Zahra Zahedi, Teruhisa Misu, Kumar Akash
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 7/8/2026, 12:03:28 PM
Summary
This paper investigates emergent inefficiencies and bottlenecks in multi-agent ad-hoc collaboration using an agent-based model set in a kitchen environment. It demonstrates that rigid specialization (the 'specialist's dilemma') and high skill assertion lead to system-level bottlenecks, workload inequality, and fragmented network silos. The study also reveals that team size and communication overhead interact with task structure (serial vs. parallel) to produce diminishing returns and redundant collaboration, highlighting the need for adaptive coordination in multi-agent system design.
Entities (9)
Relation Signals (6)
Specialist's dilemma → generates → System-level bottlenecks
confidence 98% · rigid role assertion generates system-level bottlenecks, amplifies workload inequality
Skill assertion → causes → Workload inequality
confidence 95% · Skill assertion also emerged as the primary driver of workload imbalance.
Skill assertion → leadsto → Network silos
confidence 92% · High levels of skill assertion led to high network assortativity... fragmented the network into disconnected components
Parallel task → benefitsfrom → Larger team size
confidence 88% · For the parallelizable onion soup task, larger teams and lower communication costs produced significantly more meals
Team size → interactswith → Communication overhead
confidence 88% · team size and communication overhead interact with problem structure to generate diminishing returns
Serial task → suffersfrom → Redundant collaboration
confidence 85% · larger teams... were effective for parallel tasks, but redundant for serial tasks.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Computational models of collaboration without prior coordination often overlook how heterogeneous agent traits and complex task structures jointly produce systemic bottlenecks, inefficiencies, and contribution inequalities. We address this by using an agent-based model of ad-hoc teamwork in a kitchen environment. Our model integrates diverse agent personas with tasks that combine serial and parallel dependencies. We identify a specialist's dilemma, where rigid role assertion generates system-level bottlenecks, amplifies workload inequality, and fosters fragmented, homophilous networks. We also find that team size and communication overhead interact with problem structure to generate diminishing returns and redundant collaboration. Linking micro-level behavior to macro-level outcomes provides insights into emergent collaboration and design principles for effective multi-agent teamwork.
Tags
Links
- Source: https://arxiv.org/abs/2605.08540v1
- Canonical: https://arxiv.org/abs/2605.08540v1
Trouble viewing inline? Open PDF directly →
Full Text
13,249 characters extracted from source content.
Expand or collapse full text
Too Many Specialists: Emergent Inefficiencies and Bottlenecks for Multi-agent Ad-hoc Collaboration Extended Abstract Benjamin Panny University of Pittsburgh Pittsburgh, PA, United States BMP83@pitt.edu Shashank Mehrotra Honda Research Institute USA, Inc. San Jose, CA, United States shashank_mehrotra@honda-ri.com Zahra Zahedi Honda Research Institute USA, Inc. San Jose, CA, United States zahra_zahedi@honda-ri.com Teruhisa Misu Honda Research Institute USA, Inc. San Jose, CA, United States tmisu@honda-ri.com Kumar Akash Honda Research Institute USA, Inc. San Jose, CA, United States kakash@honda-ri.com ABSTRACT Computational models of collaboration without prior coordina- tion often overlook how heterogeneous agent traits and complex task structures jointly produce systemic bottlenecks, inefficiencies, and contribution inequalities. We address this by using an agent- based model of ad-hoc teamwork in a kitchen environment. Our model integrates diverse agent personas with tasks that combine serial and parallel dependencies. We identify a specialist’s dilemma, where rigid role assertion generates system-level bottlenecks, am- plifies workload inequality, and fosters fragmented, homophilous networks. We also find that team size and communication overhead interact with problem structure to generate diminishing returns and redundant collaboration. Linking micro-level behavior to macro- level outcomes provides insights into emergent collaboration and design principles for effective multi-agent teamwork. KEYWORDS Multi-agent systems, Agent-based models, Social simulation, Col- laborative Networks ACM Reference Format: Benjamin Panny, Shashank Mehrotra, Zahra Zahedi, Teruhisa Misu, and Ku- mar Akash. 2026. Too Many Specialists: Emergent Inefficiencies and Bottle- necks for Multi-agent Ad-hoc Collaboration: Extended Abstract. In Proc. of the 25th International Conference on Autonomous Agents and Multiagent Sys- tems (AAMAS 2026), Paphos, Cyprus, May 25 – 29, 2026, IFAAMAS, 3 pages. https://doi.org/10.65109/CYXP1261 1 INTRODUCTION Collaboration requires agents to balance initiative, adaptability, and mutual awareness under uncertainty [13], particularly in ad- hoc settings where agents lack prior coordination. Computational models of teamwork can alleviate these challenges by predicting the conditions under which collaboration will succeed or fail [11, 12]. Proc. of the 25th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2026), C. Amato, L. Dennis, V. Mascardi, J. Thangarajah (eds.), May 25 – 29, 2026, Paphos, Cyprus.© 2026 International Foundation for Autonomous Agents and Multiagent Systems (w.ifaamas.org). This is the author’s version of the work. It is posted here for your personal use. Not for redistribution. The definitive Version of Record was published in Proceedings of Proc. of the 25th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2026) (AAMAS ’26). Figure 1: Kitchen Environment and Problem Structure. (A) Orders vary the ratios of parallel (Soup) vs. serial (Steak) meals. (B) Agents claiming tasks become “leaders.” (C) Serial Steak sequence. (D) Parallelizable Soup cycle. . Frameworks such as the Transactive Systems framework [4, 10,14,15] and Collective Adaptation [2] highlight how socially networked reasoning and knowledge enable team reconfiguration when environments change. However, current models rarely scale beyond small teams or account for serial-parallel task structures [6,7]. As a result, few models formally link macro-level perfor- mance, bottlenecks, and inequalities observed in real teams to micro- level agent behaviors observed in the individuals that constitute these teams. In this paper, we address this gap through an agent-based model of ad-hoc teamwork in a physically grounded kitchen environment. By introducing heterogeneous agent personas endowed with so- cial traits—such as collaboration initiative, agreeableness, and skill assertion—we show how bottlenecks, inequalities, and subgroup formations can emerge from simple local interaction rules. Our find- ings offer mechanistic levers for human-AI team design: seeding proactive leadership, mitigating rigid specialization, and bridging silos through adaptive coordination. 2 COLLABORATION MODEL We developed an agent-based model (ABM) where agents collab- orate to complete cooking tasks in a 2D grid environment (Cf. arXiv:2605.08540v1 [cs.MA] 8 May 2026 [1]). We vary team size [5], communication costs [9], and agent social/transactive factors [2, 3] (Figure 1). 2.1 Serial and Parallel Kitchen Tasks Real-world collaboration tasks are rarely perfectly divisible. We designed two recipes modeling this constraint: (1) Steak: A serial task with a fixed sequence (get meat, grill, plate, serve). (2) Onion Soup: A parallelizable task with repeated substeps (get onion, chop onion, place in stove) that multiple agents can perform concurrently. 2.2Agent Attributes Determine Agent Rule-sets Agents’ attributes result in actions via a decentralized Affordance- Context-Action (ACA) loop governed by four randomly assigned traits. Agreeableness controls the likelihood of accepting help requests, enabling tests of passive cooperation. Collaboration Ini- tiative determines whether an agent initiates interactions, allowing us to measure the impact of proactive communicators. Task Distri- bution Preference governs the choice between starting unclaimed tasks or joining existing ones, introducing differences in organi- zational strategy (i.e., complete started tasks before starting new tasks or not). Finally, Skill Assertion determines whether agents refuse tasks outside their expertise; this operationalizes the spe- cialist’s dilemma [3,8] by introducing tension between individual competence and team efficiency. 3 RESULTS 3.1 Scaling Effects and Diminishing Returns Team size and communication affected serial and parallel tasks differently. For the parallelizable onion soup task (Figure 2a), larger teams and lower communication costs produced significantly more meals, though this was subject to diminishing returns. For the serial steak task, larger teams and higher communication costs produced steaks faster than their counterparts with low communication costs (Figure 2b). This is because low communication costs promoted larger teams that were effective for parallel tasks, but redundant for serial tasks. Thus, coordination overhead can be beneficial when it restricts redundant team building. (a) Onion Soup (Parallel)(b) Steak (Serial) Figure 2: (a) Parallel tasks (Onion Soup) scale with team size and broad communication. (b) Serial tasks (Steak) benefit from higher communication cost, which results in smaller, more efficient teams. 3.2 Bottlenecks from Specialization We identified a specialist’s dilemma, where agents asserting their specific skills creates system-level bottlenecks. A high proportion (100%) of skill-asserting agents resulted in a significant decrease in total meals completed. However, this was partially rectified when specialists were also proactive communi- cators with task distribution preferences, as this persona enabled specialists to find tasks that fit their specialty earlier in the simula- tion. Skill assertion also emerged as the primary driver of workload imbalance. When agents rigidly adhered to roles (e.g., "I only grill"), they remained idle when they could have supported agents working on other tasks. 3.3 Emergent Silos Agent traits significantly altered the topology of the emergent collaboration network. High levels of skill assertion led to high network assortativity (0.578) and modularity (Figure 3a). Agents naturally formed homophilous clusters, interacting only with those sharing their specialty. This "siloing" effect fragmented the network into disconnected components, increasing the average shortest path length and reducing global integration. As skill assertion decreased, the network transformed from isolated specialist clusters into a dense, integrated collaborative web. (a) 100% Skill Assertion(b) 0% Skill Assertion Figure 3: (a) When 100% of agents assert skills, high assor- tativity and disconnected components emerge. (b) When 0% of agents assert skills, a dense, integrated collaborative web forms. 4 DISCUSSION & CONCLUSION Micro-level agent traits and task structures interact to shape the success and failure of ad-hoc teamwork. We show that rigid special- ization leads to the specialist’s dilemma—generating bottlenecks, inequality, and fragmented social silos. We also show that increas- ing communication costs can be beneficial if it reduces redundant teamwork. These findings suggest that effective MAS design must actively counteract the tendency for specialists to isolate, perhaps by algorithmically encouraging agents to occasionally accept tasks outside their primary expertise to maintain system-wide resilience. REFERENCES [1]Micah Carroll, Rohin Shah, Mark Ho, Tom Griffiths, Pieter Abbeel, and Anca Dragan. 2020. Overcooked-ai: A benchmark for multi-agent learning under partial observability. In Proceedings of the 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)(2020). 2374–2380. [2]Mirta Galesic, Daniel Barkoczi, Andrew M. Berdahl, Dora Biro, Giuseppe Car- bone, Ilaria Giannoccaro, Robert L. Goldstone, Cleotilde Gonzalez, Anne Kandler, Albert B. Kao, Rachel Kendal, Michelle Kline, Eun Lee, Giovanni Francesco Mas- sari, Alex Mesoudi, Henrik Olsson, Niccolo Pescetelli, Sabina J. Sloman, Paul E. Smaldino, and Daniel L. Stein. 2023. Beyond collective intelligence: Collective adaptation. Journal of The Royal Society Interface 20, 200 (March 2023), 20220736. https://doi.org/10.1098/rsif.2022.0736 Publisher: Royal Society. [3]Pranav Gupta. 2022. Transactive Systems Model of Collective Intelligence: The Emergence and Regulation of Collective Attention, Memory, and Reasoning. thesis. Carnegie Mellon University. https://doi.org/10.1184/R1/20039555.v1 [4]Pranav Gupta and Anita Williams Woolley. 2021. Articulating the Role of Artifi- cial Intelligence in Collective Intelligence: A Transactive Systems Framework. Proceedings of the Human Factors and Ergonomics Society Annual Meeting 65, 1 (Sept. 2021), 670–674. https://doi.org/10.1177/1071181321651354c [5]Vincent Larivière, Yves Gingras, Cassidy R. Sugimoto, and Andrew Tsou. 2015.Team size matters: Collaboration and scientific impact since 1900. Journal of the Association for Information Science and Technol- ogy 66, 7 (2015), 1323–1332.https://doi.org/10.1002/asi.23266_eprint: https://asistdl.onlinelibrary.wiley.com/doi/pdf/10.1002/asi.23266. [6] Soo Ling Lim, Peter J Bentley, Randall S Peterson, Xiaoran Hu, and JoEllyn Prouty McLaren. 2023. Kill chaos with kindness: Agreeableness improves team performance under uncertainty. Collective Intelligence 2, 1 (Jan. 2023), 26339137231158584. https://doi.org/10.1177/26339137231158584 Publisher: SAGE Publications. [7] Winter Mason and Duncan J. Watts. 2012. Collaborative learning in networks. Proceedings of the National Academy of Sciences 109, 3 (Jan. 2012), 764–769. https:// doi.org/10.1073/pnas.1110069108 Publisher: Proceedings of the National Academy of Sciences. [8]Mancur Olson. 1965. The logic of collective action: public goods and the theory of groups. Harvard Univ. Press, Cambridge, Mass. https://w.hup.harvard. edu/catalog.php?isbn=9780674537514 Number: 124 Pages: 176 tex.added-at: 2011-02-21T14:53:36.000+0100 tex.interhash: ca423582ffdb8545c0567c16786fa08e tex.intrahash: d52f789a4c4711ba434fc8106806af3b tex.timestamp: 2021-11- 04T09:43:47.000+0100. [9] Michelle O’Daniel and Alan H. Rosenstein. 2008. Professional Communica- tion and Team Collaboration. In Patient Safety and Quality: An Evidence- Based Handbook for Nurses. Agency for Healthcare Research and Quality (US). https://w.ncbi.nlm.nih.gov/books/NBK2637/ [10] Yuqing Ren and Linda Argote. 2011. Transactive memory systems 1985–2010: An integrative framework of key dimensions, antecedents, and consequences. The Academy of Management Annals 5, 1 (2011), 189–229. https://doi.org/10.1080/ 19416520.2011.590300 Place: United Kingdom Publisher: Taylor & Francis. [11]Eduardo Salas, Dana E. Sims, and C. Shawn Burke. 2005. Is there a “Big Five” in Teamwork? Small Group Research 36, 5 (Oct. 2005), 555–599. https://doi.org/10. 1177/1046496405277134 Publisher: SAGE Publications Inc. [12]Randall Spain, Michael Geden, Wookhee Min, Bradford W Mott, and James C Lester. 2019. Toward Computational Models of Team Effectiveness with Natural Language Processing.. In TTW@ AIED. 30–39. [13]Milind Tambe. 1997. Towards flexible teamwork. Journal of artificial intelligence research 7 (1997), 83–124. [14]Daniel M. Wegner. 1987. Transactive Memory: A Contemporary Analysis of the Group Mind. In Theories of Group Behavior, Brian Mullen and George R. Goethals (Eds.). Springer, New York, NY, 185–208. https://doi.org/10.1007/978-1-4612- 4634-3_9 [15]Daniel M. Wegner, Toni Giuliano, and Paula T. Hertel. 1985. Cognitive In- terdependence in Close Relationships. In Compatible and Incompatible Re- lationships, William Ickes (Ed.). Springer New York, New York, NY, 253–276. https://doi.org/10.1007/978-1-4612-5044-9_12