Paper deep dive
Efficient Failure Management for Multi-Agent Systems with Reasoning Trace Representation
Lingzhe Zhang, Tong Jia, Mingyu Wang, Weijie Hong, Chiming Duan, Minghua He, Rongqian Wang, Xi Peng, Meiling Wang, Gong Zhang, Renhai Chen, Ying Li
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 94%
Last extracted: 3/26/2026, 2:29:24 AM
Summary
EAGER is an efficient failure management framework for Multi-Agent Systems (MASs) that utilizes reasoning trace representation. It employs unsupervised reasoning-scoped contrastive learning to encode intra-agent reasoning and inter-agent coordination, enabling real-time failure detection, diagnosis, and reflexive mitigation based on historical failure patterns.
Entities (6)
Relation Signals (3)
EAGER → managesfailurein → Multi-Agent Systems
confidence 100% · EAGER, an efficient failure management framework for multi-agent systems
EAGER → usesmethod → Reasoning-Scoped Contrastive Learning
confidence 95% · EAGER employs unsupervised reasoning-scoped contrastive learning
EAGER → evaluatedon → AutoGen-Code
confidence 90% · We conduct preliminary experiments on three public open-source MAS to evaluate the effectiveness of EAGER
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Large Language Models (LLM)-based Multi-Agent Systems (MASs) have emerged as a new paradigm in software system design, increasingly demonstrating strong reasoning and collaboration capabilities. As these systems become more complex and autonomous, effective failure management is essential to ensure reliability and availability. However, existing approaches often rely on per-trace reasoning, which leads to low efficiency, and neglect historical failure patterns, limiting diagnostic accuracy. In this paper, we conduct a preliminary empirical study to demonstrate the necessity, potential, and challenges of leveraging historical failure patterns to enhance failure management in MASs. Building on this insight, we propose \textbf{EAGER}, an efficient failure management framework for multi-agent systems based on reasoning trace representation. EAGER employs unsupervised reasoning-scoped contrastive learning to encode both intra-agent reasoning and inter-agent coordination, enabling real-time step-wise failure detection, diagnosis, and reflexive mitigation guided by historical failure knowledge. Preliminary evaluations on three open-source MASs demonstrate the effectiveness of EAGER and highlight promising directions for future research in reliable multi-agent system operations.
Tags
Links
- Source: https://arxiv.org/abs/2603.21522v1
- Canonical: https://arxiv.org/abs/2603.21522v1
Trouble viewing inline? Open PDF directly →
Full Text
34,775 characters extracted from source content.
Expand or collapse full text
EAGER: Efficient Failure Management for Multi-Agent Systems with Reasoning Trace Representation Lingzhe Zhang Peking University Beijing, China zhang.lingzhe@stu.pku.edu.cn Tong Jia ∗ Peking University Beijing, China jia.tong@pku.edu.cn Mingyu Wang Peking University Beijing, China mingyuwang25@stu.pku.edu.cn Weijie Hong Peking University Beijing, China hongwj@stu.pku.edu.cn Chiming Duan Peking University Beijing, China duanchiming@stu.pku.edu.cn Minghua He Peking University Beijing, China hemh2120@stu.pku.edu.cn Rongqian Wang Huawei Technologies Co., Ltd. Beijing, China wangrongqian2@huawei.com Xi Peng Huawei Technologies Co., Ltd. Hong Kong SAR, China pancy.pengxi@huawei.com Meiling Wang Huawei Technologies Co., Ltd. Shenzhen, China wangmeiling17@huawei.com Gong Zhang Huawei Technologies Co., Ltd. Shenzhen, China nicholas.zhang@huawei.com Renhai Chen Huawei Technologies Co., Ltd. Beijing, China chenrenhai@huawei.com Ying Li ∗ Peking University Beijing, China li.ying@pku.edu.cn ABSTRACT Large Language Models (LLM)-based Multi-Agent Systems (MASs) have emerged as a new paradigm in software system design, increas- ingly demonstrating strong reasoning and collaboration capabilities. As these systems become more complex and autonomous, effective failure management is essential to ensure reliability and availability. However, existing approaches often rely on per-trace reasoning, which leads to low efficiency, and neglect historical failure patterns, limiting diagnostic accuracy. In this paper, we conduct a preliminary empirical study to demonstrate the necessity, potential, and chal- lenges of leveraging historical failure patterns to enhance failure management in MASs. Building on this insight, we propose EA- GER, an efficient failure management framework for multi-agent systems based on reasoning trace representation. EAGER employs unsupervised reasoning-scoped contrastive learning to encode both intra-agent reasoning and inter-agent coordination, enabling real- time step-wise failure detection, diagnosis, and reflexive mitigation guided by historical failure knowledge. Preliminary evaluations on three open-source MASs demonstrate the effectiveness of EAGER and highlight promising directions for future research in reliable multi-agent system operations. CCS CONCEPTS • Software and its engineering→ Maintaining software. KEYWORDS Failure Management, AgentOps, Multi Agents, Reasoning 1 INTRODUCTION Multi-Agent Systems (MASs) powered by Large Language Models (LLMs) are rapidly emerging as a new paradigm in software sys- tems design [5,6,10,15,18,25,32,38,49]. By decomposing complex tasks into coordinated agents, each capable of reasoning, planning, and interacting with tools or environments, these systems offer unprecedented flexibility and autonomy. Such MASs are increas- ingly applied across diverse domains, including software engineer- ing [11–13,39–41,46,47], intelligent assistants [14,27,30,33,45], and scientific workflows [8,9,26], demonstrating their remarkable capability in handling complex, dynamic, and multi-step problems. However, the very capabilities that make LLM-powered MASs powerful also introduce unique operational challenges: their dy- namic behavior and reasoning processes can lead to unpredictable failures, and traditional monitoring or debugging approaches are often inadequate to ensure efficient and accurate system operation. To address these challenges, the concept of Agent System Oper- ations (AgentOps) has been proposed [29], offering a systematic framework for managing, diagnosing, and mitigating failures in multi-agent systems by leveraging agent-level reasoning traces alongside traditional monitoring data. MAST [2] and TRAIL [4] analyze failures in general-purpose MASs. Ruofan et al. [23] and Simiao et al. [22] focus on failures in code-based MASs. Xuyan et al. [24] examine failures in platform- orchestrated agentic systems. Meanwhile, Alfonso et al. [1], TrustA- gent [36], G-Safeguard [28], and GUARDIAN [50] primarily ad- dress anomaly detection, which determines whether a particular MAS reasoning process has failed or been compromised. In con- trast, Who&When [48], AgenTracer [37], AEGIS [17], FAMAS [7], A2P [31], and RAFFLES [51] focus on failure diagnosis, identifying which agent (Who), at which step (When), experienced which error arXiv:2603.21522v1 [cs.SE] 23 Mar 2026 Conference’17, July 2017, Washington, DC, USALingzhe Zhang et al. (What) that led to task failure. Although the effectiveness of these methods has been demonstrated in general-purpose MASs, they encounter several challenges when applied to a practical MAS. •Per-Trace Reasoning Leads to Low Efficiency. Reasoning traces generated by MASs are rich in semantic information. Cur- rent methods typically process each trace independently, relying on a large judge LLM to perform reasoning [31,37,51] for seman- tic anomaly detection as well as failure diagnosis. While effective, this approach is extremely time-consuming: not only does it an- alyze each trace individually, but the use of a large judge LLM further increases computational overhead. In a high-throughput MAS handling numerous requests, this combination results in significantly reduced operational efficiency. •Neglecting Historical Failure Patterns Limits Accuracy. While current methods leverage large judge LLMs to infer human- like solutions for novel failures based on the rich semantics of reasoning traces, these LLMs are inherently unstable. This insta- bility means that the same failure may sometimes be analyzed correctly and sometimes incorrectly. Although methods such as RAFFLES [51] introduce a set of specialized evaluators to assess the quality of the judge’s reasoning, this merely adds layers of evaluation without addressing the underlying issue. Therefore, effectively leveraging historical failure patterns is crucial to en- sure that previously encountered failures are recognized and handled reliably, thereby improving overall accuracy. For complex general-purpose MASs, the aforementioned chal- lenges are difficult to address due to the unpredictable and highly variable nature of failures [2,4,16,19–21,43,44]. In contrast, a specific practical MAS typically exhibits relatively fixed task types, stable agent invocation chains, and a consistent organizational structure. As a result, its failure types and patterns are compara- tively limited. When AgentOps methods originally designed for general-purpose MASs are applied to such a specific system, histor- ical reasoning traces can be leveraged to enable fast and accurate failure management, avoiding redundant analysis and mitigating the inherent instability of judge LLMs. However, constructing an effective representation model for reasoning traces involves several critical problems that must be addressed: (P1) Representation Learning for Reasoning-Oriented Traces.To accelerate failure understanding, models must capture reasoning traces that fundamentally differ from conventional textual or operational representations. Each trace encodes both intra-agent reasoning dynamics and inter-agent coordination logic. Learning representations that preserve both scopes while maintaining discriminative power across diverse reasoning styles presents a significant challenge. (P2) Lack of Labeled and Generalizable Failure Data. Improving diagnostic accuracy requires semantically rich supervision. Yet, labeled multi-agent failure data—particularly those with explicit root causes—are extremely scarce. Moreover, such data often exhibit strong domain bias, limiting the generalization of trained models. Consequently, developing a learning paradigm capable of lever- aging unlabeled or weakly structured traces for reasoning-level alignment is essential. Recognizing these challenge, we first conduct a preliminary em- pirical study to gain a deeper understanding of the underlying characteristics of failures in LLM-based MAS. Specifically, we aim to demonstrate that (i) failures within a fixed multi-agent system are often limited and recurring rather than arbitrary, suggesting strong potential for historical pattern reuse; and (i) existing state- of-the-art text embedding models fail to effectively capture the reasoning structure and relational semantics inherent in agent rea- soning traces, indicating the need for a dedicated representation learning paradigm. Building on these insights, we propose EAGER — an Efficient failure management framework for multi-AGEnt systems with Reasoning trace representation. EAGER models the reasoning trace generated during multi-agent collaboration, capturing both intra- agent reasoning dynamics and inter-agent orchestration patterns. A Representation Model, trained via Reasoning-Scoped Contrastive Learning, learns to encode these two reasoning scopes into a uni- fied latent space, enabling efficient comparison and retrieval of similar reasoning experiences from historical traces. During system execution, EAGER performs Step-Wise Detection to identify po- tential failures at each reasoning stage. Once detected, it triggers a Reflexive Mitigation process, allowing agents to self-reflect, replan, or regenerate responses for rapid recovery. When a final output is confirmed as incorrect by the user, EAGER conducts Expert Inspect + Agent RCA to update both fine-grained (agent-level) and coarse- grained (system-level) failure knowledge, thereby continuously improving the system’s failure management ability. We conduct preliminary experiments on three public open- source MAS to evaluate the effectiveness of EAGER in anomaly detection and failure diagnosis. Furthermore, its instantaneous de- tection and reflexive mitigation mechanisms can also enhance re- sponse accuracy in specific task. 2 PRELIMINARY EMPIRCAL STUDY 2.1 Failure Pattern Concentration in Multi-Agent Systems To investigate how failures distribute across different reasoning pro- cesses, we analyze three open-source MASs—AutoGen-Code [23], RCLAgent [42], and SWE-Agent [35]. Using their official task datasets, we manually examine failed cases and categorize them into representative failure types. Table 1: Failure Type Analysis across Three MAS Failure TypeAutoGen-Code RCLAgent SWE-Agent Decomposition Error34.48%0.00%0.00% Incorrect Code 48.28%0.00%46.15% Round Limitation 17.24%5.26%0.00% Critical Trace Miss0.00%52.63%0.00% Metrics Query Error0.00%42.11%0.00% Editing Error0.00%0.00%25.64% Localization Error0.00%0.00%28.21% As shown in Table 1, the three MASs exhibit highly distinct fail- ure distributions. AutoGen-Code mainly fails due to Incorrect Code and Decomposition Error, RCLAgent due to Critical Trace Miss and Metrics Query Error, while SWE-Agent often suffers from Editing and Localization Errors. These results reveal a clear concentration of failure patterns within specific reasoning or coordination scopes, Efficient Failure Management for Multi-Agent Systems with Reasoning Trace RepresentationConference’17, July 2017, Washington, DC, USA User: How long to drive Paris → Berlin? Planner: 1. get_distance 2. estimate_time Calculator: receives distance = 1050 → time = 1050 mins Response: formats reply → “It will take 1050 minutes.” User: How long to drive Rome → Milan? Planner: 1. get_distance 2. estimate_time Map: get_distance (Rome, Milan) → distance = 600 Calculator: receives distance = 600 → time = 600 mins Response: formats reply → “It will take 600 minutes.” Detection: orchestration-centric similar with failure knowledge — — Mitigation: Failed! Rethink the whole orchestration. Response: “It will take 6.5 hours from Rome to Milan.” ...... User: Convert $100 to EUR Planner: 1. fetch_rate 2. compute_amount Searcher: fetch_rate (USD,EUR) → rate = 0.85 Calculator: receives rate = 0.85 → amount = 8500 Detection: model-centric similar with failure knowledge — — Mitigation: Failed! Rethink the Calculator agent. Response: “100 USD = 85 EUR” Calculator: receives rate = 0.85 → amount = 85 Map: get_distance(Paris, Berlin) → distance = 1050 — Failure Knowledge — Multi-Agent Systems Generate Inter Agent Reasoning Trace User Question Answer Reflexive Mitigation Representation Model Failure Knowledge Fine-Gained Coarse-Gained Intra Agent Step-Wise Detection Failure? Y N Wrong? Y User Expert Inspect Offline Reasoning Scoped Contrastive Learning Agent RCA + orchestration-centric model-centric Figure 1: Workflow of EAGER. The left panel shows the overall framework, while the right illustrates an example. suggesting that historical failure knowledge can be leveraged to enhance the efficiency and accuracy of MAS failure management. 2.2 Evaluation of Existing Embeddings on Reasoning Trace Representation To assess whether current state-of-the-art (SOTA) embedding models can effectively represent reasoning traces, we conduct a retrieval-based experiment using two leading models: Qwen3-0.6B- Embedding [34] and BGE-M3-Embedding [3]. Specifically, we use reasoning traces generated by RCLAgent on nine distinct questions, each queried five times, yielding a total of 45 reasoning traces. Hu- man annotators verify that the five traces corresponding to the same question are semantically similar. Table 2: Similar Reasoning Trace Retrieval Experiment ModelRecall@10 NDCG@10 MRR@10 Qwen3-0.6B-Embedding13.3%8.7%6.2% BGE-M3-Embedding22.2%14.5%10.8% As shown in Table 2, both models perform poorly in retrieving semantically similar reasoning traces. This suggests that current text embedding models—though highly effective in general NLP tasks—struggle to capture the complex hierarchical reasoning and coordination semantics inherent to multi-agent reasoning traces. These findings call for a representation method that models reason- ing scope and agent interactions. 3 METHODOLOGY Building on the findings of preliminary empirical study, we propose EAGER, an Efficient failure management framework for multi- AGEnt systems with Reasoning trace representation. As shown in Figure 1, EAGER begins by capturing reasoning traces generated during multi-agent interactions, which contain both intra-agent reasoning processes and inter-agent orchestration patterns. A ded- icated Representation Model, trained via Reasoning-Scoped Con- trastive Learning, learns to encode these traces into a unified latent space that aligns both intra-agent and inter-agent semantics. During system operation, EAGER performs Step-Wise Detection, which continuously evaluates each reasoning step against accumu- lated historical Failure Knowledge to identify potential failures in real time. Upon detecting a failure, a Reflexive Mitigation mechanism is triggered—allowing agents to self-reflect, replan, or regenerate responses—thereby effectively recovering from most failures. Finally, when the multi-agent system produces an output that the user deems incorrect, EAGER triggers an optional Expert In- spect + Agent RCA procedure. This process first employs SOTA Conference’17, July 2017, Washington, DC, USALingzhe Zhang et al. Multi-Agent Systems Questions Agent-1 Agent-2 Agent-n ...... Trace Encoder Agent Questions Question Variation Positive PairsNegative Pairs Semantic Aligned Progressive Aligned Reasoning Consistent Action Consistent In-Batch Negatives Temporal Negatives Cross-Agent Negatives Reasoning Encoder Reasoning-1 Reasoning-2 Reasoning-n ... ... ... ++ Intra-scope Contrastive Loss Inter-scope Contrastive Loss Prefix→Full Ranking Loss Training Data ConstructionDual-EncoderLoss Objectives Figure 2: Training Pipeline of Reasoning Scoped Contrsative Learning AgentOps methods to analyze the corresponding reasoning trace, after which expert judgment is applied to validate and refine the findings. The resulting insights are used to enrich both fine-grained (agent-level) and coarse-grained (system-level) failure knowledge, which is then incorporated into the historical knowledge base to continually improve future detection and mitigation. 3.1 Failure Management Components Failure Knowledge. The failure knowledge is composed of both fine-grained knowledge and coarse-grained knowledge. Fine- grained knowledge refers to the identification of specific reasoning errors in an individual agent’s reasoning process. For example, it captures the specific reasoning steps or decisions within the agent’s workflow where hallucinations, inconsistencies, or other errors occur. This knowledge is more granular and helps pinpoint the source of the problem at the agent level. Coarse-grained knowledge, on the other hand, refers to an understanding of the overall failure at the trace level, where the entire reasoning trace is recognized as erroneous, but the exact agent or step responsible for the failure re- mains undetermined. This level of knowledge is crucial for quickly detecting failures without knowing the specific source. Step-Wise Detection. By leveraging failure knowledge for matching, anomaly detection becomes significantly faster. There- fore, EAGER performs step-wise detection, conducting a check after each agent completes its reasoning. Specifically, the reason- ing embedding generated by the current agent is matched with the fine-grained knowledge. Once all agents have completed their reasoning, the entire reasoning trace is embedded and compared with the coarse-grained knowledge. If any similarity is detected, a fault is identified, triggering the reflexive mitigation process. Reflexive Mitigation. Most non-malicious failures in multi- agent systems can be self-corrected through reflective reasoning. Accordingly, EAGER incorporates a reflexive mitigation mechanism that operates at two complementary levels. When step-wise detec- tion precisely identifies a specific agent’s failure, EAGER performs a model-centric reflection to refine the reasoning process of that agent. Conversely, when the entire reasoning trace is deemed faulty, EAGER triggers an orchestration-centric reflection, enabling the system to re-evaluate inter-agent coordination and restore consis- tent reasoning dynamics at the global level. 3.2 Reasoning Scoped Contrsative Learning The training pipeline of reasoning-scoped contrastive learning is illustrated in Figure 2. To ensure a highly generalizable representa- tion model, we intentionally avoid relying on failure-labeled data. Instead, we construct training samples through question variation, under the hypothesis that semantically similar questions tend to yield reasoning traces with analogous structures and logical pro- gressions in most cases. From this process, we extract both system-level reasoning traces (across all agents) and agent-level reasoning segments (from in- dividual agents), which naturally provide positive and negative pairs required for contrastive learning. We then jointly train two hierarchical encoders: the Reasoning Encoder, which captures intra- agent reasoning semantics, and the Trace Encoder, which integrates multiple reasoning embeddings to encode inter-agent orchestration dependencies. Ltotal= 휆 1 Lintra+ 휆 2 Linter+ 휆 3 Lrank,(1) The overall objective functionL total is composed of three com- plementary components, formulated as Equation 1, whereL intra represents the intra-scope contrastive loss, enforcing proximity between reasoning embeddings of the same agent under similar question variations,L inter denotes the inter-scope contrastive loss, aligning reasoning patterns across agents to preserve coor- dination semantics, andL rank is the prefix-to-full ranking loss, which encourages consistency between partial and complete rea- soning traces, enabling robust representation of reasoning trace. 4 PRELIMINARY EVALUATION To evaluate the feasibility and effectiveness of EAGER, we conduct experiments on the three aforementioned open-source MASs. The results for anomaly detection and failure diagnosis, evaluated by F1-Score, as well as the average detection latency, are presented in Table 3. EAGER demonstrates consistently strong performance Efficient Failure Management for Multi-Agent Systems with Reasoning Trace RepresentationConference’17, July 2017, Washington, DC, USA Table 3: Anomaly Detection & Diagnosis Results TaskAutoGen-Code RCLAgent SWE-Agent Anomaly Detection73.57%86.18%79.95% Failure Diagnosis63.23%78.76%69.51% Detection Latency5.23s4.57s4.91s across all tasks, highlighting its potential for efficient and accurate failure management in multi-agent systems. However, our current representation model is only lightly fine-tuned from the Qwen-0.6B- Embedding backbone, which limits its generalization capability. In future work, we plan to perform large-scale fine-tuning and further optimization to enhance both cross-domain robustness and reasoning adaptability. Table 4: Task Performance Improvement with EAGER MethodR1R3R5 R10 MRR RCLAgent28.47% 62.37% 64.41% 68.14% 46.13% RCLAgent + EAGER30.19% 65.82% 68.56% 70.03% 48.65% Furthermore, since EAGER enables instantaneous detection and reflexive mitigation, it can also enhance the response accuracy of specific tasks. To verify this, we conducted additional experi- ments on RCLAgent, where task performance was evaluated using Recall@푘(R@푘) and Mean Reciprocal Rank (MRR). As shown in Table 4, integrating EAGER with RCLAgent yields consistent and noticeable improvements across all evaluation metrics, demonstrat- ing its effectiveness in enhancing real-time reasoning reliability. 5 CONCLUSION This paper explores the necessity, potential, and challenges of lever- aging historical failure patterns to enhance failure management in MASs. Building on these insights, we propose EAGER, an efficient failure management framework for MASs with reasoning-trace rep- resentation. Preliminary experiments demonstrate the feasibility and effectiveness of EAGER. In future work, we plan to further im- prove its generalization capability through large-scale fine-tuning. ACKNOWLEDGMENT This work was supported by the Huawei–Peking University Joint Laboratory of Mathematics. REFERENCES [1]Alfonso Amayuelas, Xianjun Yang, Antonis Antoniades, Wenyue Hua, Liang- ming Pan, and William Yang Wang. 2024. MultiAgent Collaboration Attack: Investigating Adversarial Attacks in Large Language Model Collaborations via Debate. In Findings of the Association for Computational Linguistics: EMNLP 2024. 6929–6948. [2]Mert Cemri, Melissa Z Pan, Shuyi Yang, Lakshya A Agrawal, Bhavya Chopra, Rishabh Tiwari, Kurt Keutzer, Aditya Parameswaran, Dan Klein, Kannan Ram- chandran, et al.2025. Why do multi-agent llm systems fail? arXiv preprint arXiv:2503.13657 (2025). [3]Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. arXiv preprint arXiv:2402.03216 (2024). [4]Darshan Deshpande, Varun Gangal, Hersh Mehta, Jitin Krishnan, Anand Kan- nappan, and Rebecca Qian. 2025. TRAIL: Trace Reasoning and Agentic Issue Localization. arXiv preprint arXiv:2505.08638 (2025). [5]Zhuoyun Du, Chen Qian, Wei Liu, Zihao Xie, Yifei Wang, Yufan Dang, Weize Chen, and Cheng Yang. 2024. Multi-agent software development through cross- team collaboration. arXiv preprint arXiv:2406.08979 (2024). [6] Chiming Duan, Minghua He, Pei Xiao, Tong Jia, Xin Zhang, Zhewei Zhong, Xiang Luo, Yan Niu, Lingzhe Zhang, Yifan Wu, et al.2025. LogAction: Consistent Cross- system Anomaly Detection through Logs via Active Domain. arXiv preprint arXiv:2510.03288 (2025). [7]Yu Ge, Linna Xie, Zhong Li, Yu Pei, and Tian Zhang. 2025. Who is introducing the failure? automatically attributing failures of multi-agent systems via spectrum analysis. arXiv preprint arXiv:2509.13782 (2025). [8]Alireza Ghafarollahi and Markus J Buehler. 2025. SciAgents: automating scientific discovery through bioinspired multi-agent intelligent graph reasoning. Advanced Materials 37, 22 (2025), 2413523. [9]Juraj Gottweis, Wei-Hung Weng, Alexander Daryin, Tao Tu, Anil Palepu, Petar Sirkovic, Artiom Myaskovsky, Felix Weissenberger, Keran Rong, Ryutaro Tanno, et al. 2025. Towards an AI co-scientist. arXiv preprint arXiv:2502.18864 (2025). [10]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 Methodology 34, 5 (2025), 1–30. [11]Minghua He, Chiming Duan, Pei Xiao, Tong Jia, Siyu Yu, Lingzhe Zhang, Weijie Hong, Jin Han, Yifan Wu, Ying Li, et al.2025. United we stand: Towards end-to- end log-based fault diagnosis via interactive multi-task learning. arXiv preprint arXiv:2509.24364 (2025). [12]Minghua He, Tong Jia, Chiming Duan, Pei Xiao, Lingzhe Zhang, Kangjin Wang, Yi- fan Wu, Ying Li, and Gang Huang. 2025. Walk the talk: Is your log-based software reliability maintenance system really reliable? arXiv preprint arXiv:2509.24352 (2025). [13] Weijie Hong, Yifan Wu, Lingzhe Zhang, Chiming Duan, Pei Xiao, Minghua He, Xixuan Yang, and Ying Li. 2025. CSLParser: A Collaborative Framework Using Small and Large Language Models for Log Parsing. In 2025 IEEE 36th International Symposium on Software Reliability Engineering (ISSRE). IEEE, 61–72. [14] Mengkang Hu, Yuhang Zhou, Wendong Fan, Yuzhou Nie, Bowei Xia, Tao Sun, Ziyu Ye, Zhaoxuan Jin, Yingru Li, Qiguang Chen, et al.2025. Owl: Optimized workforce learning for general multi-agent assistance in real-world task automa- tion. arXiv preprint arXiv:2505.23885 (2025). [15]Yue Hu, Yuzhu Cai, Yaxin Du, Xinyu Zhu, Xiangrui Liu, Zijie Yu, Yuchen Hou, Shuo Tang, and Siheng Chen. [n. d.]. Self-Evolving Multi-Agent Collaboration Networks for Software Development. In The Thirteenth International Conference on Learning Representations. [16]Xiaosong Huang, Hongyi Liu, Yifan Wu, Lingzhe Zhang, Tong Jia, Ying Li, and Zhonghai Wu. 2025. UDA-RCL: Unsupervised Domain Adaptation for Microser- vice Root Cause Localization Utilizing Multimodal Data. IEEE Transactions on Services Computing (2025). [17] Fanqi Kong, Ruijie Zhang, Huaxiao Yin, Guibin Zhang, Xiaofei Zhang, Ziang Chen, Zhaowei Zhang, Xiaoyuan Zhang, Song-Chun Zhu, and Xue Feng. 2025. AEGIS: Automated Error Generation and Identification for Multi-Agent Systems. arXiv preprint arXiv:2509.14295 (2025). [18]Han Li, Yuling Shi, Shaoxin Lin, Xiaodong Gu, Heng Lian, Xin Wang, Yantao Jia, Tao Huang, and Qianxiang Wang. 2025. Swe-debate: Competitive multi-agent debate for software issue resolution. arXiv preprint arXiv:2507.23348 (2025). [19] Hongyi Liu, Xiaosong Huang, Mengxi Jia, Tong Jia, Jing Han, Zhonghai Wu, and Ying Li. 2024. Uac-ad: Unsupervised adversarial contrastive learning for anomaly detection on multi-modal data in microservice systems. IEEE Transactions on Services Computing 17, 6 (2024), 3887–3900. [20]Hongyi Liu, Xiaosong Huang, Mengxi Jia, Lingzhe Zhang, Tong Jia, Zhonghai Wu, and Ying Li. 2025. AAAD: Asynchronous Inter-Variable Relationship-Aware Anomaly Detection for Multivariate Time Series. In 2025 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 1–6. [21]Hongyi Liu, Yinping Ma, Xiaosong Huang, Lingzhe Zhang, Tong Jia, and Ying Li. 2025. ORA: Job Runtime Prediction for High-Performance Computing Platforms Using the Online Retrieval-Augmented Language Model. In Proceedings of the 39th ACM International Conference on Supercomputing. 884–894. [22]Simiao Liu, Fang Liu, Liehao Li, Xin Tan, Yinghao Zhu, Xiaoli Lian, and Li Zhang. 2025. An Empirical Study on Failures in Automated Issue Solving. arXiv preprint arXiv:2509.13941 (2025). [23]Ruofan Lu, Yichen Li, and Yintong Huo. 2025. Exploring Autonomous Agents: A Closer Look at Why They Fail When Completing Tasks. arXiv preprint arXiv:2508.13143 (2025). [24] Xuyan Ma, Xiaofei Xie, Yawen Wang, Junjie Wang, Boyu Wu, Mingyang Li, and Qing Wang. 2025.Diagnosing Failure Root Causes in Platform- Orchestrated Agentic Systems: Dataset, Taxonomy, and Benchmark. arXiv preprint arXiv:2509.23735 (2025). [25] Leyi Pan, Zheyu Fu, Yunpeng Zhai, Shuchang Tao, Sheng Guan, Shiyu Huang, Lingzhe Zhang, Zhaoyang Liu, Bolin Ding, Felix Henry, et al.2025. Omni- SafetyBench: A Benchmark for Safety Evaluation of Audio-Visual Large Language Models. arXiv preprint arXiv:2508.07173 (2025). [26]Kyle Swanson, Wesley Wu, Nash L Bulaong, John E Pak, and James Zou. 2024. The virtual lab: Ai agents design new sars-cov-2 nanobodies with experimental validation. bioRxiv (2024), 2024–11. Conference’17, July 2017, Washington, DC, USALingzhe Zhang et al. [27]Junyang Wang, Haiyang Xu, Haitao Jia, Xi Zhang, Ming Yan, Weizhou Shen, Ji Zhang, Fei Huang, and Jitao Sang. 2024. Mobile-agent-v2: Mobile device operation assistant with effective navigation via multi-agent collaboration. Advances in Neural Information Processing Systems 37 (2024), 2686–2710. [28]Shilong Wang, Guibin Zhang, Miao Yu, Guancheng Wan, Fanci Meng, Chongye Guo, Kun Wang, and Yang Wang. 2025. G-safeguard: A topology-guided se- curity lens and treatment on llm-based multi-agent systems. arXiv preprint arXiv:2502.11127 (2025). [29] Zexin Wang, Jingjing Li, Quan Zhou, Haotian Si, Yuanhao Liu, Jianhui Li, Gaogang Xie, Fei Sun, Dan Pei, and Changhua Pei. 2025. A Survey on AgentOps: Cate- gorization, Challenges, and Future Directions. arXiv preprint arXiv:2508.02121 (2025). [30] Di Wen, Kunyu Peng, Junwei Zheng, Yufan Chen, Yitain Shi, Jiale Wei, Ruiping Liu, Kailun Yang, and Rainer Stiefelhagen. 2025. Mica: Multi-agent industrial coordination assistant. arXiv preprint arXiv:2509.15237 (2025). [31]Alva West, Yixuan Weng, Minjun Zhu, Zhen Lin, Zhiyuan Ning, and Yue Zhang. 2025. Abduct, Act, Predict: Scaffolding Causal Inference for Automated Failure Attribution in Multi-Agent Systems. arXiv preprint arXiv:2509.10401 (2025). [32]Pei Xiao, Tong Jia, Chiming Duan, Minghua He, Weijie Hong, Xixuan Yang, Yihan Wu, Ying Li, and Gang Huang. 2025. CLSLog: Collaborating Large and Small Models for Log-based Anomaly Detection. In Proceedings of the 33rd ACM International Conference on the Foundations of Software Engineering. 686–690. [33]Zhenran Xu, Xue Yang, Yiyu Wang, Qingli Hu, Zijiao Wu, Longyue Wang, Weihua Luo, Kaifu Zhang, Baotian Hu, and Min Zhang. 2025. ComfyUI-Copilot: An Intelligent Assistant for Automated Workflow Development. arXiv preprint arXiv:2506.05010 (2025). [34]An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al.2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388 (2025). [35] John Yang, Carlos E Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. 2024. Swe-agent: Agent-computer interfaces enable automated software engineering. Advances in Neural Information Processing Systems 37 (2024), 50528–50652. [36]Miao Yu, Fanci Meng, Xinyun Zhou, Shilong Wang, Junyuan Mao, Linsey Pan, Tianlong Chen, Kun Wang, Xinfeng Li, Yongfeng Zhang, et al.2025. A survey on trustworthy llm agents: Threats and countermeasures. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 6216–6226. [37] Guibin Zhang, Junhao Wang, Junjie Chen, Wangchunshu Zhou, Kun Wang, and Shuicheng Yan. 2025. AgenTracer: Who Is Inducing Failure in the LLM Agentic Systems? arXiv preprint arXiv:2509.03312 (2025). [38]Lingzhe Zhang, Liancheng Fang, Chiming Duan, Minghua He, Leyi Pan, Pei Xiao, Shiyu Huang, Yunpeng Zhai, Xuming Hu, Philip S Yu, et al.2025. A survey on parallel text generation: From parallel decoding to diffusion language models. arXiv preprint arXiv:2508.08712 (2025). [39] Lingzhe Zhang, Tong Jia, Mengxi Jia, Yifan Wu, Aiwei Liu, Yong Yang, Zhonghai Wu, Xuming Hu, Philip Yu, and Ying Li. 2025. A Survey of AIOps in the Era of Large Language Models. Comput. Surveys (2025). [40]Lingzhe Zhang, Tong Jia, Mengxi Jia, Yifan Wu, Hongyi Liu, and Ying Li. 2025. ScalaLog: Scalable Log-Based Failure Diagnosis Using LLM. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 1–5. [41]Lingzhe Zhang, Tong Jia, Mengxi Jia, Yifan Wu, Hongyi Liu, and Ying Li. 2025. XRAGLog: A Resource-Efficient and Context-Aware Log-Based Anomaly Detec- tion Method Using Retrieval-Augmented Generation. In AAAI 2025 Workshop on Preventing and Detecting LLM Misinformation (PDLM). [42]Lingzhe Zhang, Tong Jia, Kangjin Wang, Weijie Hong, Chiming Duan, Minghua He, and Ying Li. 2025. Adaptive root cause localization for microservice systems with multi-agent recursion-of-thought. arXiv preprint arXiv:2508.20370 (2025). [43] Lingzhe Zhang, Tong Jia, Yunpeng Zhai, Leyi Pan, Chiming Duan, Minghua He, Mengxi Jia, and Ying Li. 2026. Agentic Memory Enhanced Recursive Reasoning for Root Cause Localization in Microservices. arXiv preprint arXiv:2601.02732 (2026). [44]Lingzhe Zhang, Tong Jia, Yunpeng Zhai, Leyi Pan, Chiming Duan, Minghua He, Pei Xiao, and Ying Li. 2026. Hypothesize-Then-Verify: Speculative Root Cause Analysis for Microservices with Pathwise Parallelism. arXiv preprint arXiv:2601.02736 (2026). [45]Lingzhe Zhang, Yunpeng Zhai, Tong Jia, Chiming Duan, Minghua He, Leyi Pan, Zhaoyang Liu, Bolin Ding, and Ying Li. 2025. MicroRemed: Benchmarking LLMs in Microservices Remediation. arXiv preprint arXiv:2511.01166 (2025). [46]Lingzhe Zhang, Yunpeng Zhai, Tong Jia, Chiming Duan, Siyu Yu, Jinyang Gao, Bolin Ding, Zhonghai Wu, and Ying Li. 2025. ThinkFL: Self-Refining Failure Localization for Microservice Systems via Reinforcement Fine-Tuning. arXiv preprint arXiv:2504.18776 (2025). [47] Lingzhe Zhang, Yunpeng Zhai, Tong Jia, Xiaosong Huang, Chiming Duan, and Ying Li. 2025. Agentfm: Role-aware failure management for distributed databases with llm-driven multi-agents. In Proceedings of the 33rd ACM International Con- ference on the Foundations of Software Engineering. 525–529. [48]Shaokun Zhang, Ming Yin, Jieyu Zhang, Jiale Liu, Zhiguang Han, Jingyang Zhang, Beibin Li, Chi Wang, Huazheng Wang, Yiran Chen, et al.[n. d.]. Which Agent Causes Task Failures and When? On Automated Failure Attribution of LLM Multi-Agent Systems. In Forty-second International Conference on Machine Learning. [49]Yiran Zhang, Ruiyin Li, Peng Liang, Weisong Sun, and Yang Liu. 2025. Knowledge- based multi-agent framework for automated software architecture design. In Proceedings of the 33rd ACM International Conference on the Foundations of Soft- ware Engineering. 530–534. [50]Jialong Zhou, Lichao Wang, and Xiao Yang. 2025. GUARDIAN: Safeguarding LLM Multi-Agent Collaborations with Temporal Graph Modeling. arXiv preprint arXiv:2505.19234 (2025). [51]Chenyang Zhu, Spencer Hong, Jingyu Wu, Kushal Chawla, Charlotte Tang, Youb- ing Yin, Nathan Wolfe, Erin Babinsky, and Daben Liu. 2025. RAFFLES: Reasoning- based Attribution of Faults for LLM Systems. arXiv preprint arXiv:2509.06822 (2025).