Paper deep dive
[COMP25] The Automated Negotiating Agents Competition (ANAC) 2025 Challenges and Results
Reyhan Aydoğan, Tim Baarslag, Tamara C. P. Florijn, Katsuhide Fujita, Catholijn M. Jonker, Yasser Mohammad
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 4/16/2026, 1:09:15 AM
Summary
The paper presents the results and challenges of the 15th International Automated Negotiating Agents Competition (ANAC 2025), held at IJCAI 2025. It details two primary leagues: the Automated Negotiation League (ANL), focusing on sequential multi-deal negotiations, and the Supply Chain Management League (SCML), focusing on concurrent negotiations in market simulations. The study highlights the trade-offs between computational scalability and strategic foresight, the effectiveness of domain-specific heuristics over complex modeling, and the introduction of human-agent negotiation benchmarks.
Entities (7)
Relation Signals (4)
ANAC 2025 → heldat → IJCAI 2025
confidence 100% · one of the official competitions of IJCAI 2025.
ANAC 2025 → utilized → NegMAS
confidence 100% · Both leagues utilized the NegMAS [ Mohammad et al., 2021 ] framework
RUFL → won → ANL 2025
confidence 95% · One of the winner agents, RUFL
AS0 → won → SCML-Standard
confidence 95% · The winner of SCML-Standard (AS0 developed by Atsunaga et al.)
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:This paper presents the primary research challenges and key findings from the 15th International Automated Negotiating Agents Competition (ANAC 2025), one of the official competitions of IJCAI 2025. We focus on two critical domains: multi-deal negotiations and the development of agents capable of concurrent negotiation within complex supply chain management environments. Furthermore, this work analyzes the results of the competition and outlines strategic directions for future iterations.
Tags
Links
- Source: https://arxiv.org/abs/2604.13914v1
- Canonical: https://arxiv.org/abs/2604.13914v1
Trouble viewing inline? Open PDF directly →
Full Text
17,572 characters extracted from source content.
Expand or collapse full text
[COMP25] The Automated Negotiating Agents Competition (ANAC) 2025 Challenges and Results Reyhan Aydo ̆ gan 1,2 , Tim Baarslag 3,4 , Tamara C.P. Florijn 3,5 , Katsuhide Fujita 6 , Catholijn M. Jonker 2,7 , Yasser Mohammad 8 1 ̈ Ozye ̆ gin ̈ University, 2 Delft University of Technology, 3 Centrum Wiskunde & Informatica (CWI), 4 Eindhoven University of Technology, 5 Utrecht University, 6 Tokyo University of Agriculture and Technology & National Institute for Advanced Industrial Science and Technology, 7 Leiden University, 8 NEC Corporation & National Institute for Advanced Industrial Science and Technology reyhan.aydogan@ozyegin.edu.tr, T.Baarslag@cwi.nl, tamara.florijn@cwi.nl, katfuji@c.tuat.ac.jp, C.M.Jonker@tudelft.nl, y.mohammad@nec.com Abstract This paper presents the primary research challenges and key findings from the 15th International Au- tomated Negotiating Agents Competition (ANAC 2025), one of the official competitions of IJCAI 2025. We focus on two critical domains: multi- deal negotiations and the development of agents capable of concurrent negotiation within complex supply chain management environments. Further- more, this work analyzes the results of the com- petition and outlines strategic directions for future iterations. 1 Introduction Since 2010, the International Automated Negotiating Agents Competition (ANAC) has spearheaded innovation in au- tonomous agent research, introducing foundational chal- lenges to the multi-agent systems community [ Jonker et al., 2017 ] . Since its inception with 7 participants [ Baarslag et al., 2012 ] , the competition has expanded significantly, surpass- ing 850 total participants across 16 editions [ Baarslag et al., 2015 ] . The 2025 competition featured two distinct leagues. The Automated Negotiation League (ANL) required partici- pants to develop bilateral agents that can engage in sequential multi-deal negotiations with multiple opponents and inter- dependent utility functions. In the Supply Chain Manage- ment League (SCML) [ Mohammad et al., 2019 ] , the objec- tive is to design agents that maximize profit within a competi- tive market. These agents must strategically navigate concur- rent negotiations to secure raw materials and sell manufac- tured goods. The 2025 edition attracted 142 international researchers forming 42 teams, competing for a 2,500 EUR prize pool. Both leagues utilized the NegMAS [ Mohammad et al., 2021 ] framework for agent development and tournament execution. Agents competed in a round-robin format across randomized scenarios. ANL performance was evaluated via individual utility and Nash distance, whereas SCML rankings were de- termined by cumulative profit within the market simulation. The subsequent sections detail the specific league configura- tions and analyze the final results. Moreover, ANAC 2025 introduced a pilot competition be- tween IJCAI 2025 participants in which these participants negotiated against autonomous agents in a variety of scenar- ios for a 1,000 EUR prize pool. The scenarios included a trade negotiation inspired by the Supply Chain Management League, an Island survival tool distribution negotiation and a Grocery items distribution scenario. The participants con- ducted 1, 456 negotiations and two winners were announced at the closing session of IJCAI 2025. 2 Main League Challenges & Results In the main league of 2025, each participating agent encoun- ters multiple opponents in sequence and is rewarded for the specific combination of the deals made in each negotiation. The challenge of this league is sequential multi-deal negoti- ation, where the participants need to coordinate an enormous number of options available and design an agent that outper- forms other contestants by conceding efficiently and obtain- ing the best deals. At the start of each negotiation, one agent will receive the role center agent, while the others receive the role edge agent. The center agent negotiates against multiple opponents (the edge agents) in a sequential manner, in what we call subnego- tiations. Each subnegotiation will follow the Alternating Of- fers Protocol [ Aydo ̆ gan et al., 2014 ] where the starting agent makes an opening offer, which is followed by acceptance, a counteroffer, or a walk-away, repeated in a turn-taking fash- ion. If the negotiation ends, the center agent continues to the subsequent subnegotiation. The agents negotiate over special types of domains that are characterized by the utility function of the center agent, two of which we discuss here. In the Job Hunt scenario, a job hunter (center agent) negotiates about two issues: the num- ber of office days and the salary. In the end, the center utility function is defined in such a way that he receives the maxi- mum of all the deals that he gathered. His preferences within a subnegotiation (called side utility function) are modeled us- ing a linear additive function. Secondly, the Target Quantity scenario models a buyer (center agent) that negotiates with arXiv:2604.13914v1 [cs.MA] 15 Apr 2026 multiple sellers over the purchase quantity of a single prod- uct. The buyer aims for a specific quantity, for example to buy 10 products i.e. a peak preference at 10 products, while buying more or less gradually decreases the utility to 0. In total, 17 teams submitted to ANL 2025, of which 12 were selected as finalists. The results of this year are shown in Table 1. As the negotiation protocol is sequential, the agents need to take into account what future actions will influence their current best decision. When the number of bids and thus the number of possible future actions grows, the number of combined outcomes to consider grows exponentially. One of the winner agents, RUFL, tackled the computa- tional challenge by constructing a tree to estimate the ex- pected utility. The agent maintains a probability distribution over all the expected values of the child outcomes where their probabilities are generated via a softmax over their expected values with some temperature. By limiting the tree search at a certain level, they keep their computational time within lim- its. The second winner agent, SAC Agent, is built using the Soft Actor-Critic (SAC) reinforcement learning framework. The SAC bidding policy is a time-dependent concession strat- egy, where the concession rate is guided by the SAC model. Table 1: ANL 2025 Results Rank AgentAs Center As Edge Final Score 1RUFL0.7140.0840.399 1SAC Agent0.7330.0640.399 3UFunAtAgent 0.6860.0780.382 Other participants approach the uncertain future deals and their influence on the search space in distinct ways: • Pessimistic. These agents pretend that no further deals will be made (e.g., UfunATagent, SAC, default ANL 2025 strategies). • Contingent. These agents use a probability distribution to assess the likelihood of further deals in the future (e.g., ProbaBot, RUFL, RivAgent, WAgent). • Optimistic. These agents assume that the future will un- fold exactly as planned (e.g., WAgent under condition- als). Agents using the pessimistic methods circumvent the com- putational search challenge by ignoring future deals, disre- garding any influence of future deals, limiting their perfor- mance. However, using contingent or optimistic approaches takes more computational time. All in all, the biggest chal- lenge turned out to be the memory explosion caused by the big combined outcome spaces (e.g., CARC2025, OzUA- gent, KDY, SmartNegotiator, ProbaBot), tackled with differ- ent techniques such as a dynamic target (e.g., EOHAgent, CARCAgent2025), sampling methods (e.g., the Memorizer, kAgent), dynamic programming (Astrat3m) or reinforment learning techniques (e.g., SacAgent). 3 Supply-Chain Management League The 2025 Supply Chain Management League (SCML) com- prised two tracks: OneShot and Standard. A total of 20 qual- ified teams participated, with 11 competing in the OneShot track and 9 in the Standard track. Because the game mechan- ics remained identical to SCML 2024, participants were able to leverage the open-source code and technical reports of pre- vious agents. To further support the research community, the 2025 edi- tion simplified reinforcement learning and MARL integration by providing a standard Gymnasium [ Towers et al., 2024 ] and a petting-zoo [ Terry et al., 2021 ] environments that encapsu- late the SCML simulation [ Mohammad et al., 2025 ] . The primary objective of SCML is to align automated ne- gotiation research with the practical challenges of industrial applications. The simulation serves as an abstraction of a fun- damental procurement dilemma: How can production needs be met while minimizing the combined costs of procurement and inventory? In the SCML OneShot world, multiple au- tonomous agents manage factories within a supply chain, buying raw materials and selling final products. The market is driven by a set of exogenous un-negotiable contracts for the raw materials and final products. Every simulated days, agents must reach agreements with their suppliers/consumers for buying their production needs and selling their produced items. Contracts specify price and quantity (and delivery date in case of SCML-Standard); failure to meet agreed-upon obli- gations results in penalties. The ultimate goal for all agents is to maximize their accumulated profit. SCML’s two tracks offer different levels of complexity re- garding the negotiation search space and temporal dependen- cies: • OneShot Track: Products are perishable, meaning prof- its or losses on a given day are independent of other days, except in the indirect effect of changing partners’ future negotiation behavior. Agents act as either buyers or sellers (never both), and the search space is limited by small price and quantity ranges. The research focus here is on repeated concurrent negotiation in many-to-many environments. • Standard Track: This track introduces higher complex- ity by making products nonperishable (carrying storage costs) and allowing agents to negotiate delivery dates. The production graph is deeper, requiring agents to man- age both buying and selling simultaneously. This shifts the challenge toward designing strategies for dependent sequential sets of concurrent negotiations. Table 2: SCML 2025 Results TrackRank AgentScore OnShotCautiousOneShotAgent1.0915 1CostAverseAgent1.0896 1Rchan1.0895 1AlmostEqualAgent1.0892 Standard1AS01.009 PenguinAgent0.992 2XenoSotaAgent0.960 3Ultra Super Miracle Final Agent Z 0.937 The performance of the participants is summarized in Ta- ble 2. Notably, the winners of both tracks prioritized domain- specific heuristics over complex opponent modeling (for the second year in a row). Since 2020, all finalists of SCML could outperform the winner of the previous year. This pat- tern started to break in 2025 with the winner of SCML- OneShot 2024 (CautiousOneshotAgent) remaining the top performing agent in 2025 and the winner of SCML-Standard 2024 (PenguinAgent) only outperformed by the winner of SCML-Standard 2025 (AS0). Three agents tied for the winner of SCML-Oneshot: 1) CostAverseAgent (developed by Yuzuru Kitamura) which fo- cuses on minimizing the risk of financial loss by prioritizing cost-based decision-making over aggressive profit-seeking. 2) Rchan (developed by Shota Takayama) which is built around a sophisticated dynamic aspiration model and mar- ket estimation. 3) AlmostEqualAgent (developed by Kaito Miwa) which utilizes a strategy centered on proportional eq- uity and collaborative surplus sharing to reach agreements quickly. The winner of SCML-Standard (AS0 developed by At- sunaga et al.) focuses on a risk-balanced production-first strategy that adapts to the complexities of a full supply chain simulation. It employs a conservative negotiation strategy that sets price floors using expected manufacturing costs and storage fees, ensuring that every contract contributes to a pos- itive net profit. 4 Conclusion The 16th International Automated Negotiating Agents Com- petition (ANAC 2025) has once again demonstrated the evolving complexity of autonomous negotiation in multi- agent systems. By shifting the focus toward sequential multi- deal negotiations in ANL and deepened supply chain depen- dencies in the SCML, the competition continues to push the boundaries of how agents handle high-dimensional outcome spaces and temporal uncertainty. Our analysis of the 2025 results yields several key insights: • Computational Scalability: In ANL, the transition to multi-deal scenarios highlighted a critical trade-off be- tween computational overhead and strategic foresight. Success was largely defined by an agent’s ability to manage ”memory explosion” through tree-search prun- ing, sampling, or reinforcement learning (e.g., the SAC Agent). • Heuristics vs. Complexity: In SCML, the continued success of domain-specific heuristics and risk-averse strategies (e.g., AS0 and CostAverseAgent) suggests that in highly volatile, concurrent markets, robustness and cost-containment often outperform complex oppo- nent modeling. • Human-Agent Interaction: The 2025 pilot competi- tion at IJCAI provided a rare benchmark for how au- tonomous agents perform against human negotiators in survival and trade scenarios, highlighting the need for more ”explainable” and human-centric negotiation strategies. Looking forward, the competition aims to further bridge the gap between theoretical frameworks and industrial ap- plication.Future iterations of ANAC will likely explore the integration of Large Language Models (LLMs) for more nuanced communication in the Human-Agent Negotiation League planned for 2026. ANL will explore the design of a negotiation agent for bilateral negotiation that tries to mis- lead its opponent. The agent is rewarded for the agreement made in the negotiation, as well as for how well it is able to deceive its opponent. This negotiation-in-the-wild challenge will explore the role of deception and its counter-measures in automated negotiation. As usual with ANAC since 2010, all agent code is open- sourced and is accessible from the submission website at https://anac.cs.brown.edu. As agent-based negotiation moves closer to real-world ap- plications, the lessons learned from ANAC 2025 provide a vital roadmap for developing resilient, efficient, and collabo- rative autonomous systems. Ethical Statement The Human-Agent Negotiation pilot design received the ap- proval of Ozyegin University, Turkey. Personal information of participants were not kept beyond the announcement of winners. Acknowledgments This publication is part of the Vidi project COMBINE (VI.Vidi.203.044) (partly) financed by the Dutch Research Council (NWO). ANAC 2024 was sponsored by AI Journal, NWO, NEC-AIST AI Cooperative Research Laboratory, and Springer AI. References [ Aydo ̆ gan et al., 2014 ] Reyhan Aydo ̆ gan, Koen V. Hindriks, and Catholijn M. Jonker. Multilateral mediated negotiation protocols with feedback. In Novel Insights in Agent-based Complex Automated Negotiation, pages 43–59. Springer Japan, Tokyo, 2014. [ Baarslag et al., 2012 ] TimBaarslag,KoenHindriks, Catholijn M. Jonker, Sarit Kraus, and Raz Lin.The first automated negotiating agents competition (ANAC 2010). In Takayuki Ito, Minjie Zhang, Valentin Robu, Shaheen Fatima, and Tokuro Matsuo, editors, New Trends in Agent-based Complex Automated Negotiations, Series of Studies in Computational Intelligence, pages 113–135, Berlin, Heidelberg, 2012. Springer-Verlag. [ Baarslag et al., 2015 ] Tim Baarslag,Reyhan Aydo ̆ gan, Koen V. Hindriks, Katsuhide Fuijita, Takayuki Ito, and Catholijn M. Jonker. The automated negotiating agents competition, 2010-2015. AI Magazine, 36(4):115–118, 12/2015 2015. [ Jonker et al., 2017 ] Catholijn M Jonker, Reyhan Aydogan, Tim Baarslag, Katsuhide Fujita, Takayuki Ito, and Koen V Hindriks.Automated negotiating agents competition (anac). In AAAI, pages 5070–5072, 2017. [ Mohammad et al., 2019 ] YasserMohammad,En- rique Areyan Viqueira, Nahum Alvarez Ayerza, Amy Greenwald, Shinji Nakadai, and Satoshi Morinaga. Supply chain management world.In Matteo Baldoni, Mehdi Dastani, Beishui Liao, Yuko Sakurai, and Rym Zalila Wenkstern, editors, PRIMA 2019: Principles and Practice of Multi-Agent Systems, pages 153–169, Cham, 2019. Springer International Publishing. [ Mohammad et al., 2021 ] YasserMohammad,Shinji Nakadai, and Amy Greenwald. Negmas: a platform for situated negotiations. In Recent Advances in Agent-based Negotiation: Formal Models and Human Aspects, pages 57–75. Springer, 2021. [ Mohammad et al., 2025 ] YasserMohammad,Shinji Nakadai, and Amy Greenwald. Automated negotiation in supply chains a generalist environment for rl/marl research.In Ryuta Arisaka, Victor Sanchez-Anguix, Sebastian Stein, Reyhan Aydo ̆ gan, Leon van der Torre, and Takayuki Ito, editors, PRIMA 2024: Principles and Practice of Multi-Agent Systems, pages 19–24, Cham, 2025. Springer Nature Switzerland. [ Terry et al., 2021 ] J Terry, Benjamin Black, Nathaniel Grammel, Mario Jayakumar, Ananth Hari, Ryan Sullivan, Luis S Santos, Clemens Dieffendahl, Caroline Horsch, Ro- drigo Perez-Vicente, et al. Pettingzoo: Gym for multi- agent reinforcement learning. Advances in Neural Infor- mation Processing Systems, 34:15032–15043, 2021. [ Towers et al., 2024 ] Mark Towers, Ariel Kwiatkowski, Jor- dan Terry, John U Balis, Gianluca De Cola, Tristan Deleu, Manuel Goul ̃ ao, Andreas Kallinteris, Markus Krimmel, Arjun KG, et al.Gymnasium: A standard interface for reinforcement learning environments. arXiv preprint arXiv:2407.17032, 2024.