cs.SEOct 11, 2026

When Messages Go Wrong: Understanding Communication Faults in LLM-Based Multi-Agent Systems

Authors: Jinhao Zhou, Xiaoyu Zhang, Jiongchi Yu, Fahao Chen, Jun Wu

Organizations: Waseda University, Japan · The Hong Kong University of Science and Technology, Hong Kong SAR, China · Nanyang Technological University, Singapore · Shandong University, China · 

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

LLM-based multi-agent systems (LLM-MAS) use specialized agents that exchange messages to accomplish complex tasks. Their reliance on networked protocols, remote tools, and model endpoints exposes communication to message loss, delay, and context expiration. Because agents often exchange open-ended natural language, anomalies can alter task-relevant information while leaving messages processable, allowing semantic errors to propagate through subsequent interactions. Systematic studies of faults in inter-agent communication remain limited. This paper presents a large-scale empirical study of how communication faults manifest, propagate, and affect task outcomes in LLM-MAS. We construct a cross-layer taxonomy linking lower-stack disturbances to application-visible faults and develop a fault-injection and tracing pipeline. Our evaluation covers more than 29,000 executions across 100 tasks, five domains, three workflow topologies, and three model configurations. We find that 20.7% of failures with complete detection records occur without an immediate failure signal, reaching 76.2% for partial-content faults. Flat maintains 53.0% task success under communication faults, compared with 3.3% and 3.2% for Sequential and Hierarchical, respectively. This advantage primarily arises when an alternative information path preserves the evidence required by downstream agents. We also evaluate execution-side and evidence protection mechanisms and find that they recover different failure cases. Evidence recovery depends on whether trustworthy information remains accessible through an unaffected path. These observations provide empirical insights into communication-fault propagation, workflow resilience, and protection design for LLM-MAS.

Figures & tables

Figure 1. A silent partial-content communication failure in an LLM-MAS. Although the message remains syntactically valid, the Verifier accepts incomplete evidence without an immediate failure signal.
Figure 2. Cross-layer communication-fault propagation. P/D/N/T denote conventional lower-stack communication layers; A represents application-level inter-agent communication and receiver-visible communication faults; M and O represent downstream MAS consequences and task outcomes.
LayerFault / effect nameBrief descriptionRepresentative source(s)
PP-F1 Bit-level corruptionIsolated or burst errors in communicated bits.( Stone and Partridge, 2000 )
 P-F2 Link/endpoint interruptionNIC, cable, or power disruption removes a path.( Braden, 1989 ; Linux Kernel Documentation, 2011 )
 P-F3 Physical-medium interferenceInterference degrades reception and delivery timing.( Karn et al., 2004 )
 P-F4 Timing instabilityClock drift affects communication timing decisions.( Mills et al., 2010 )
 P-F5 Communication-state corruptionHardware errors affect communication buffers or runtime state.( Carreira et al., 1998 ; Linux Kernel Documentation, 2026 )
DD-F1 Frame loss/local dropFrames are not delivered across a local link.( Fairhurst and Wood, 2002 ; Karn et al., 2004 )
Table 1. Cross-layer taxonomy and representative sources. P/D/N/T entries are lower-stack communication-source families; A entries are receiver-visible communication-fault families; and M entries are downstream MAS consequence groups. Representative sources for the underlying phenomena are listed in the final column.
Figure 3. AutoGen workflows and injection boundaries used in the confirmation matrix. Sequential uses one evidence path, Flat adds an alternate evidence route, and Hierarchical aggregates evidence through a supervisor. The injection boundaries are action requests, observations, and evidence handoffs.
Figure 4. Cross-layer manifestation and failure visibility under communication faults. The left panel shows the distribution of primary application-level manifestations across 3,600 faulted source-mapping executions. The right panel shows silent-failure rates among 8,335 failures with complete detection records. Lower-stack fault sources exhibit many-to-many mappings to application-level symptoms, while silent failures concentrate in partial-content and semantic-corruption faults.
Figure 5. Representative silent failure under partial-content delivery. The message remains parseable after required evidence is removed, and the Verifier accepts it without recorded detection or recovery.
ModelTopologyI2 (%)I3 (%)I4 (%)
DeepSeekSequential90.772.20.0
Hierarchical94.662.50.0
Flat87.075.994.4
Qwen3.5-9BSequential72.475.90.0
Hierarchical75.075.00.0
Flat75.078.696.4
Table 2. Matched task-success rates under non-delivery at I2, I3, and I4 in Admin. Each model uses the same matched task-repeat cohort across the three communication stages.
Figure 6. Representative impact of non-delivery at different communication stages. At I2, request loss prevents the Worker from producing the required evidence. At I4, the evidence already exists and remains available through the Flat topology’s additional path.
Figure 7. Task-failure rates under semantic corruption and partial-content faults across five task contexts with the Flat topology.
Figure 8. Impacts of topology and model configuration under communication faults. (a) Clean-conditional loss across three topologies and three model configurations. Error bars report task-clustered 95% confidence intervals. (b) Paired success advantage of Flat over Sequential and Hierarchical under five fault conditions. Flat gains concentrate at the evidence handoff stage, where its branch carries the required evidence.
Figure 9. Representative topology-dependent outcome under the same I4 non-delivery fault. Sequential and Hierarchical lose their only downstream evidence path, whereas Flat preserves the required evidence through its additional direct branch.
Figure 10. Guarded evidence recovery repairs incomplete evidence without repeating the completed action.
Figure 11. Success rates under the main protection study. Execution-side protection primarily recovers execution faults, whereas evidence protection recovers information faults.
Fault scopeSequentialFlat
Initial handoff0→180\to 1818→1818\to 18
Handoff and recovery deliveries0→00\to 018→1818\to 18
All evidence paths0→00\to 00→00\to 0
Table 3. Recovery boundary under repeated fault exposure. Entries show Baseline →\rightarrow Combined successful runs.
More in this topic

Multi-Agent LLM Systems

Multi-agent large language model systems coordinate multiple AI agents to solve tasks through communication, specialization, shared tools, and memory.

See all 1,032 papers

Explore similar work

CardsList
Oct 23, 2026

Understanding Communication Failures in Multi-Agent LLM Systems

Multi-agent LLM systems aim to improve reasoning through debate, evidence sharing, and mutual correction, but interaction can also introduce failures that final accuracy alone cannot explain. Agents may converge too early, ignore dissent, suppress correct minority answers, or preserve a wrong framing across rounds. We present a corpus-grounded taxonomy of communication failures, built from 3,211 failed multi-round trajectories across medical QA, coding, finance, multi-hop reasoning, ethics, and embodied text tasks. Using this taxonomy, we analyze failure distributions across datasets and system designs, round-level error dynamics, recovery hardness, majority transitions, minority-signal suppression, and self-reported confidence. We find that failures are dominated by recovery/correction and consensus/decision problems rather than isolated first-round mistakes, with many emerging after agents first observe peer responses. Based on these findings, we recommend stronger disagreement handling, evidence verification, minority-signal preservation, and recovery checks before finalization.

Priya Pitre, Jun Hou, Yi Fang +2

Virginia Tech, Blacksburg, VA 24061, USA

Multi-Agent LLM SystemsMulti-Agent CommunicationLLM Agent ReliabilityFailure Attribution for LLM AgentsMulti-Agent Consensus
AskPDF
May 19, 2026

What Do Agents Communicate? Characterizing Information Exchange in Multi-Agent Systems

Large Language Models (LLMs) have enabled collaborative Multi-Agent (MA) systems, where interacting agents improve performance through diverse reasoning and iterative refinement. However, these systems remain vulnerable to error propagation, where early-stage information degrades downstream reasoning. To address this, we conduct a systematic analysis of inter-agent communication to identify which information drives MA performance. We find that the absence of reasoning and verification in inter-agent communication significantly degrades performance. Based on these insights, we propose Category-Aware Recovery Augmentation (technique), which enforces the presence of critical information during communication. recovers up to 86.2% of failed cases. Our results highlight the key role of information quality in effective MA collaboration. Our code is available at https://anonymous.4open.science/r/cara_mas

Yong Jin Chun, Iftekhar Ahmed

Department of Informatics University of California, Irvine Irvine, CA, 92618

Multi-Agent CollaborationMulti-Agent SystemsMulti-Agent ReasoningMulti-Agent CommunicationLLM Agent Reliability
AskPDF
Feb 23, 2026

MAS-FIRE: Fault Injection and Reliability Evaluation for LLM-Based Multi-Agent Systems

As LLM-based Multi-Agent Systems (MAS) are increasingly deployed for complex tasks, ensuring their reliability has become a pressing challenge. Since MAS coordinate through unstructured natural language rather than rigid protocols, they are prone to semantic failures (e.g., hallucinations, misinterpreted instructions, and reasoning drift) that propagate silently without raising runtime exceptions. Prevailing evaluation approaches, which measure only end-to-end task success, offer limited insight into how these failures arise or how effectively agents recover from them. To bridge this gap, we propose MAS-FIRE, a systematic framework for fault injection and reliability evaluation of MAS. We define a taxonomy of 15 fault types covering intra-agent cognitive errors and inter-agent coordination failures, and inject them via three non-invasive mechanisms: prompt modification, response rewriting, and message routing manipulation. Applying MAS-FIRE to three representative MAS architectures, we uncover a rich set of fault-tolerant behaviors that we organize into four tiers: mechanism, rule, prompt, and reasoning. This tiered view enables fine-grained diagnosis of where and why systems succeed or fail. Our findings reveal that stronger foundation models do not uniformly improve robustness. We further show that architectural topology plays an equally decisive role, with iterative, closed-loop designs neutralizing over 40% of faults that cause catastrophic collapse in linear workflows. MAS-FIRE provides the process-level observability and actionable guidance needed to systematically improve multi-agent systems.

Jin Jia, Zhiling Deng, Zhuangbin Chen +2

Sun Yat-sen University Zhuhai, Guangdong, China

Multi-Agent LLM SystemsLLM Agent ReliabilityAI Agent EvaluationAgent Failure Analysis
AskPDF