MA-Mesh: Resilient Multi-Agent Collaboration via Role-based Self-Healing¶
Conference: EMNLP 2026
Paper: ACL Anthology
Area: LLM Reasoning
Keywords: Multi-Agent System, System Resilience, Chaos Engineering, Service Mesh, Cross-Agent Reflection, Role-based Failover
TL;DR¶
Drawing inspiration from cloud-native service meshes, this paper proposes MA-Mesh, a framework that cleanly decouples task-level collaboration logic from system-level resilience policies. By overlaying sidecar-style resilient proxies equipped with cross-agent reflection and role-based failover onto multi-agent topologies, MA-Mesh establishes an infrastructure-level self-healing layer that outperforms the competitive baseline C+I by an average of 32.51 percentage points in accuracy under diverse system chaos scenarios.
Background & Motivation¶
Multi-agent systems (MAS) driven by large language models have emerged as the dominant paradigm for solving complex, long-horizon tasks across software engineering, multifaceted tool usage, and scientific reasoning. By decomposing complex workflows into modular sub-tasks and assigning domain-specialized personas to distinct agents, collaborative networks such as MetaGPT, RTADev, AgentVerse, and EvoMAC substantially surpass individual models. However, existing multi-agent coordination architectures are fundamentally designed under nominal operating conditions. They assume that agents consistently follow prescribed communication topologies, that network and infrastructure nodes never experience hardware or transport failures, and that intermediate textual artifacts remain free from corruptions, hallucinations, or prompt drift.
When deployed in real-world distributed settings, multi-agent pipelines are acutely vulnerable to operational uncertainty and cascading errors. Hardware faults, network timeouts, orchestration anomalies, and malicious prompt injections trigger localized degradation that quickly propagates across dependent downstream agents. Existing mitigation strategies typically rely on ad-hoc, post-hoc output correction. For instance, frameworks like C+I deploy external challenger and inspector agents to audit and rewrite flawed messages. However, these centralized, correction-based mechanisms lack granular awareness of inter-agent role contracts, frequently over-detect benign deviations, and completely collapse when an agent encounters a non-responsive execution timeout or an abrupt process crash, leaving the entire workflow stalled unless subjected to an expensive, global re-orchestration.
In distributed cloud computing, large-scale microservice architectures overcome similar unreliability by adopting a service mesh, which strictly decouples the control plane (declarative policy management) from the data plane (sidecar proxies intercepting inter-service traffic to transparently execute retries, circuit breaking, and failover). Inspired by this principle, the authors adapt service mesh concepts to agentic collaboration. Core idea: decouple task execution logic from resilience policies by introducing MA-Mesh, an architecture that compiles resilience-aware specifications into executable policy code, attaching sidecar resilient proxies to agents in the data plane to dynamically detect role degradation via cross-agent reflection and recover from execution failure via role-based failover without global re-orchestration.
Method¶
Overall Architecture¶
MA-Mesh operates through a two-stage construction pipeline and a dual-plane runtime architecture. During orchestration, an LLM orchestrator (\(\phi_{\text{LLM}}\)) ingests a high-level task \(T\) and synthesizes a structured resilience-aware system specification, which contains the task's collaboration logic (sub-task decomposition, coordination patterns, and agent profiles) paired with cross-agent reflection policies. The orchestrator then translates this specification into executable Python policy code utilizing an AutoGen-based interface, binding each agent to an instantiated resilient proxy. At runtime, the architecture bifurcates cleanly into a control plane and a data plane: the static specification along with a dedicated Meta-Agent reside in the control plane to oversee dynamic reconfigurations, while task-executing agents and their sidecar resilient proxies operate in the data plane. The proxies intercept all inter-agent communications to enforce timeout monitoring, execute peer reflections, and trigger localized failover whenever degradation thresholds are breached.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Complex Long-Horizon Task T"] --> B["Resilience-Aware System Specification<br/>Sub-tasks · Coordination Patterns · Reflection Policies"]
B --> C["Multi-Agent Policy Code Synthesis<br/>AutoGen Interface · Parameterized Proxy Wrapping"]
C --> D["Control-Data Plane Decoupled Architecture<br/>Meta-Agent Control Plane + Resilient Proxy Data Plane"]
D --> E["Cross-Agent Reflection Mechanism<br/>Role-Specific Peer Evaluations Across Communication Edges"]
E --> F["Role-Based Failover Mechanism<br/>Timeout Tracking + Max Retry Triggering Meta-Agent Replacement"]
F -->|Verified Output / Replaced Agent Ready| G["Resilient Self-Healed Deliverable"]
Key Designs¶
1. Resilience-Aware System Specification: Explicitly Decoupling Business Collaboration Logic from Resilience Contracts
Existing multi-agent frameworks embed communication protocols and error-handling implicitly within ad-hoc system prompts, obscuring failure boundaries. MA-Mesh formalizes this setup by generating a structured specification \(\phi_{\text{LLM}}: T \to \{(\tau_k, c_k, P_k)\}_{k=1}^K, R\). The collaboration logic partitions the global objective into \(K\) sub-tasks \(\tau_k\), selects an optimal coordination pattern \(c_k\) from a curated structural pool (Sequential, Parallel & Merge, Debate, Refine), and delineates precise agent profiles \(P_k\) (specifying responsibilities, permitted tools, input schemas, and expected artifact structures). Simultaneously, the specification defines the cross-agent reflection topology \(R\), which formalizes reflection sources, targets, and role-specific evaluation criteria. This decoupled representation can be generated dynamically for novel tasks or superimposed onto existing static architectures (such as MetaGPT and RTADev) without altering their underlying business logic.
2. Infrastructure-as-Code Synthesis and Plane Separation: Sidecar Resilient Proxies as Non-Intrusive Runtime Interceptors
Embedding extensive self-healing and reflection rules directly into an agent's reasoning prompt severely restricts its effective context window and impairs domain reasoning. MA-Mesh adopts an Infrastructure-as-Code (IaC) design, wrapping agents into a standardized AgentWithResilientProxy class. In the resulting communication graph \(\Pi = (V, V_r, E)\), direct agent-to-agent edges are eliminated; all inter-agent traffic is rerouted through resilient proxies \(V_r\) such that \(E \subseteq (V_r \times V_r) \cup (V_r \times V) \cup (V \times V_r)\). This establishes a clean architectural boundary: the Meta-Agent and specification govern policy management in the control plane, while task-executing agents and sidecar proxies handle data transmission and transparent monitoring in the data plane. Proxies enforce execution timeouts, maintain retry counters, and buffer messages, providing robust system-level fault tolerance with minimal computational overhead.
3. Cross-Agent Reflection Mechanism: Role-Aware Degradation Detection Grounded in Peer Expectations
Conventional self-reflection is frequently confounded by the model's internal blind spots, whereas monolithic external verifiers lack the specialized context needed to evaluate domain-specific contracts. MA-Mesh addresses this via cross-agent reflection, where downstream consumer agents or collaborative peers inspect an agent's output through the lens of their own operational requirements. For example, when an Architect produces a system design in a software engineering pipeline, the Backend Developer's proxy audits the artifact for API route definitions and database schema completeness, while the Frontend Developer's proxy concurrently checks for component identifiers and template bindings. If the proxy identifies deviations from the reflection criteria, it injects structured diagnostic feedback back into the agent's context for regeneration. If an agent is not targeted by any peer, its proxy gracefully defaults to self-reflection, preventing corrupted or incomplete artifacts from propagating downstream.
4. Role-Based Failover Mechanism: Localized Role Re-Engineering to Preserve Task Continuity
When an agent crashes entirely, experiences infinite loops, or repeatedly fails reflection criteria, output-level revision becomes futile. MA-Mesh equips resilient proxies with a two-tiered failover trigger: (i) a strict response timeout threshold, and (ii) an upper bound on failed reflection attempts (Max Reflection Attempts). When either condition is met, the proxy initiates failover by invoking the control plane's Meta-Agent via role_based_failover(P). Rather than simply cloning an identical, flawed instance, the Meta-Agent analyzes the failure mode, synthesizes a refined role prompt with heightened constraints (e.g., explicit schema boundaries), and instantiates a substitute agent. The proxy transfers only the original task inputs while flushing corrupted intermediate states. This surgical, localized substitution preserves overall execution continuity, obviating the catastrophic overhead of restarting the entire multi-agent graph from scratch.
Benchmark & Evaluation Protocol¶
- Benchmarks & Datasets: Evaluations are conducted across four benchmarks in three domains: (i) Software Development, featuring expanded long-horizon Web Dev. (20 tasks, 113% longer prompts and 59% more functional units than rSDE-Bench, with automated Selenium test suites) and GUI Dev. (21 tasks, 1,613% longer prompts than FSD-Bench); (ii) Tool Use, using GAIA (requiring coordinated execution across web browsing, document parsing, multimodal analysis, and Python execution); and (iii) Scientific Reasoning, utilizing GPQA-Diamond (graduate-level multiple-choice questions in physics, chemistry, and biology).
- Chaos Engineering Perturbations: Following chaos engineering paradigms, four system-level uncertainty regimes are evaluated: Agent Chaos (abrupt termination of agents with a 30% default probability, varied between 20%–60%); Prompt Chaos (corrupting agent prompts via word shuffling, 60% deletion, or 60% alphanumeric replacement with a 20% probability); I/O Chaos (applying identical perturbations to message payloads during inter-agent transmission); and Compound Chaos (concurrently injecting all three chaos modes). Supplementary evaluations include adversarial prompt/output injection attacks and noisy tool environments.
- Evaluation Metrics & Models: Performance is evaluated via task Accuracy (Acc, test case pass rate for software development, exact match for QA) and Detection Recall (DR, the ratio of correctly identified affected agents). The default model across all components and baselines is
gpt-4.1-mini-0414(temperature 0.0), with reasoning-oriented models (o4-mini,gemini-3-flash, andgpt-5.4) explored in ablation studies.
Key Experimental Results¶
Main Results¶
The comparative performance between MA-Mesh and prominent baselines (single-agent CoT; static frameworks MetaGPT, RTADev, LLM-Debate; autonomous systems AgentVerse, EvoMAC; and resilient system C+I) across nominal conditions and the four chaos scenarios is summarized below (quoted directly from Table 1, averaged across three runs):
| Benchmark | Method | Vanilla Acc (%) | Agent Chaos Acc (%) / DR | Prompt Chaos Acc (%) / DR | I/O Chaos Acc (%) / DR | Compound Chaos Acc (%) / DR |
|---|---|---|---|---|---|---|
| GUI Dev. | CoT | 66.02±2.53 | 0.00±0.00 / 0.00 | 25.03±5.87 / 0.00 | 0.00±0.00 / 0.00 | 3.96±4.40 / 0.00 |
| RTADev | 66.30±1.60 | 21.87±8.35 / 0.61 | 39.64±3.27 / 0.42 | 33.90±13.38 / 0.19 | 14.99±8.29 / 0.43 | |
| MetaGPT | 70.51±4.22 | 7.39±6.10 / 0.00 | 21.62±5.12 / 0.00 | 22.63±11.46 / 0.00 | 0.00±0.00 / 0.00 | |
| C+I | 68.05±2.46 | 4.46±4.47 / 0.24 | 19.30±6.23 / 0.29 | 47.37±4.24 / 0.86 | 1.59±2.75 / 0.30 | |
| AgentVerse | 66.53±6.59 | 32.87±6.03 / 0.00 | 44.61±8.08 / 0.00 | 37.89±12.94 / 0.00 | 26.36±6.13 / 0.00 | |
| EvoMAC | 72.79±1.96 | 48.01±11.28 / 0.00 | 52.15±8.34 / 0.00 | 63.65±7.45 / 0.00 | 24.38±11.92 / 0.00 | |
| MA-Mesh | 75.29±4.34 | 71.05±1.64 / 1.00 | 65.30±5.48 / 0.84 | 67.87±5.07 / 0.82 | 65.66±5.69 / 0.90 | |
| Web Dev. | CoT | 45.73±3.02 | 0.00±0.00 / 0.00 | 0.60±1.04 / 0.00 | 0.00±0.00 / 0.00 | 0.19±0.32 / 0.00 |
| RTADev | 40.17±6.78 | 12.68±4.60 / 0.44 | 8.85±0.53 / 0.46 | 16.74±4.50 / 0.20 | 2.68±4.12 / 0.36 | |
| MetaGPT | 42.09±7.03 | 4.81±2.77 / 0.00 | 6.30±5.03 / 0.00 | 11.91±3.30 / 0.00 | 0.41±0.35 / 0.00 | |
| C+I | 45.25±3.67 | 5.72±2.81 / 0.10 | 12.47±9.17 / 0.31 | 14.03±0.39 / 0.75 | 0.00±0.00 / 0.25 | |
| AgentVerse | 52.01±2.06 | 11.95±6.38 / 0.00 | 21.55±8.06 / 0.00 | 12.50±1.38 / 0.00 | 18.38±3.77 / 0.00 | |
| EvoMAC | 41.33±4.47 | 20.45±4.72 / 0.00 | 24.29±2.56 / 0.00 | 14.72±7.17 / 0.00 | 5.80±5.10 / 0.00 | |
| MA-Mesh | 59.68±1.84 | 53.98±1.55 / 1.00 | 53.11±1.82 / 0.94 | 54.12±1.58 / 0.75 | 49.76±0.32 / 0.91 | |
| GAIA | CoT | 28.61±3.57 | 0.00±0.00 / 0.00 | 1.01±0.35 / 0.00 | 4.85±1.60 / 0.00 | 0.61±0.61 / 0.00 |
| LLM-Debate | 32.32±1.85 | 13.13±2.45 / 0.00 | 19.39±2.10 / 0.00 | 15.96±3.45 / 0.00 | 5.25±0.93 / 0.00 | |
| C+I | 30.10±0.92 | 14.75±3.11 / 0.53 | 20.81±1.95 / 0.62 | 17.58±5.97 / 0.99 | 3.70±1.91 / 0.72 | |
| AgentVerse | 24.31±2.38 | 13.18±1.45 / 0.00 | 7.07±1.75 / 0.00 | 11.31±5.44 / 0.00 | 7.68±6.48 / 0.00 | |
| MA-Mesh | 32.73±1.60 | 29.90±4.94 / 1.00 | 26.46±3.55 / 0.83 | 29.09±0.61 / 0.92 | 24.44±4.59 / 0.94 | |
| GPQA-Diamond | CoT | 64.14±1.01 | 0.00±0.00 / 0.00 | 27.94±0.58 / 0.00 | 30.97±2.10 / 0.00 | 20.04±1.54 / 0.00 |
| LLM-Debate | 66.50±1.27 | 26.77±7.02 / 0.00 | 35.52±7.65 / 0.00 | 41.41±4.01 / 0.00 | 15.32±3.44 / 0.00 | |
| C+I | 65.15±1.33 | 35.35±3.81 / 0.66 | 52.36±2.04 / 0.52 | 51.67±4.22 / 0.96 | 16.33±2.49 / 0.66 | |
| AgentVerse | 62.46±2.59 | 28.62±2.96 / 0.00 | 31.14±3.04 / 0.00 | 27.95±11.21 / 0.00 | 22.90±6.11 / 0.00 | |
| MA-Mesh | 67.68±2.20 | 61.87±1.07 / 1.00 | 62.46±3.79 / 0.81 | 61.11±2.53 / 0.98 | 61.45±3.04 / 0.96 |
Ablation Study¶
1. Component Decomposition of Resilient Proxy (Web Dev., from Table 7 & Table 19):
| Configuration | Vanilla Acc (%) | Agent Chaos Acc (%) / DR | Prompt Chaos Acc (%) / DR | I/O Chaos Acc (%) / DR | Compound Chaos Acc (%) / DR |
|---|---|---|---|---|---|
| MA-Mesh Full Framework | 59.68 | 53.98 / 1.00 | 53.11 / 0.94 | 54.12 / 0.75 | 49.76 / 0.91 |
| – w/o Cross-agent Reflection | 49.89 | 52.27 / 1.00 | 3.13 / 0.00 | 9.75 / 0.00 | 17.57 / 0.57 |
| – w/o Role-based Failover | 52.50 | 3.50 / 1.00 | 14.79 / 0.93 | 7.00 / 0.95 | 4.13 / 0.35 |
| – w/o Full Resilient Proxy | 46.06 | 0.00 / 0.00 | 1.61 / 0.00 | 8.89 / 0.00 | 0.00 / 0.00 |
2. Augmenting Static Baselines with MA-Mesh (GUI Dev., from Table 5):
| Baseline Setup | Agent Chaos Acc (%) / DR | Prompt Chaos Acc (%) / DR | I/O Chaos Acc (%) / DR |
|---|---|---|---|
| MetaGPT | 7.39 / 0.00 | 21.62 / 0.00 | 22.63 / 0.00 |
| MetaGPT + MA-Mesh | 66.86 / 1.00 (+59.47) | 56.90 / 0.88 (+35.28) | 59.09 / 0.89 (+36.46) |
| RTADev | 21.87 / 0.61 | 39.64 / 0.42 | 33.90 / 0.19 |
| RTADev + MA-Mesh | 62.13 / 1.00 (+40.26) | 56.88 / 0.75 (+17.24) | 64.65 / 0.91 (+30.75) |
3. API Cost and Wall-Clock Latency Analysis (Web Dev., from Table 6 & Table 14):
| Method | Agent Cost ($) | Resilient Cost ($) | Total Cost ($) | Agent Time (min) | Resilient Time (min) | Total Time (min) |
|---|---|---|---|---|---|---|
| RTADev | $0.180 | – | $0.180 | 16.26 | – | 16.26 |
| C+I | $0.049 | $0.155 | $0.204 | 5.12 | 3.14 | 8.26 |
| MA-Mesh | $0.114 | $0.055 | $0.170 | 3.76 | 0.77 | 4.53 |
Key Findings¶
- Asymmetric Self-Healing Superiority Under Severe Perturbations: When subjected to Agent Chaos and Compound Chaos, all existing baselines exhibit near-total functional collapse. Under Compound Chaos, C+I drops to 1.59% on GUI Dev. and 0.00% on Web Dev., whereas MA-Mesh sustains robust accuracy of 65.66% and 49.76%, respectively. Because C+I relies on textual post-editing, it cannot revive deadlocked or crashed agents; in contrast, MA-Mesh's role-based failover detects timeouts with 100% recall and initializes cleanly re-scoped replacement agents.
- Strict Functional Complementarity of Reflection and Failover: The ablation data confirms that neither self-healing mechanism suffices in isolation. Removing role-based failover causes accuracy under Agent Chaos to collapse from 53.98% to 3.50%. Removing cross-agent reflection causes accuracy under Prompt Chaos and I/O Chaos to plummet from 53%+ down to 3.13% and 9.75%, demonstrating that reflection handles semantic/logical degradation while failover safeguards infrastructural availability.
- Cost Reduction via Concurrency and Targeted Review: Despite introducing dedicated proxy layers, MA-Mesh's resilience component incurs less than half the API cost of C+I ($0.055 vs. \(0.155), and its total cost is lower than that of undefended RTADev (\)0.170 vs. $0.180). Furthermore, because the orchestrator synthesizes Parallel & Merge topologies for independent sub-tasks (accounting for 23% of structures in software tasks), MA-Mesh slashes execution latency to 4.53 minutes—less than one-third that of RTADev (16.26 min) and nearly half that of C+I (8.26 min).
Highlights & Insights¶
- Systematic Mapping of Cloud Microservice Principles to Agentic Systems: The conceptual port of service mesh principles (control vs. data plane separation and the sidecar pattern) cleanly addresses the dilemma of embedding operational fault tolerance directly into task-level reasoning prompts.
- Role Refinement Outperforms Naive Re-Execution: As shown in Table 18, under persistent failure modes such as DoS prompt injection, simple retry-upon-failure recovers only 13.36% accuracy on Web Dev., whereas MA-Mesh achieves 57.51%. The Meta-Agent dynamically reconstructs role prompts with explicit constraints, proving that architectural role adaptation fundamentally surpasses rote retry loops.
- Empirical Validation on Production-Grade Cloud Infrastructure: Beyond simulated text perturbations, MA-Mesh is validated on AIOpsLab, deploying agents as individual Kubernetes Pods subjected to real CPU/memory stress, disk I/O latency, and pod-to-pod network packet loss via Chaos Mesh (Table 9). MA-Mesh consistently maintains 55%–60% accuracy, demonstrating direct practical applicability to enterprise deployments.
Limitations & Future Work¶
- Static Allocation of Proxy Reasoning Models: While deploying advanced reasoning models (e.g., o4-mini, gpt-5.4) within resilient proxies markedly improves detection recall and task accuracy (Tables 8 & 15), model assignment remains static. Future extensions could implement dynamic multi-tier routing to invoke expensive reasoning models only upon initial anomaly detection.
- Inherent Trust in Control Plane Availability: MA-Mesh assumes that the control plane specification and the Meta-Agent remain continuously accessible and fault-free throughout the execution cycle. Hardened defenses against adversarial control plane tampering or catastrophic cluster controller failures were not explored.
- Loss of Intermediate State During Failover: The current failover mechanism resets the agent's historical conversational context to prevent corruptions from lingering, passing only raw input specifications. While this ensures semantic purity, developing selective state-checkpointing and memory-splicing mechanisms could accelerate recovery in exceptionally deep workflows.
Related Work & Insights¶
- vs C+I (Huang et al., ICML 2025): C+I implements an ad-hoc correction scheme using global challenger and inspector agents, which cannot handle physical node crashes and suffers from false-positive over-correction; MA-Mesh introduces an infrastructure-level sidecar proxy layer that pairs fine-grained role reflection with dynamic failover, delivering substantial accuracy gains under compound chaos.
- vs RTADev (Liu et al., ACL Findings 2025): RTADev uses reviewer agents for inter-stage intention alignment checks within a fixed software engineering lifecycle; MA-Mesh abstracts resilience into declarative specifications that can wrap any arbitrary topology, augmenting RTADev with failover capabilities and improving its resilience by 29.42 pp on average.
- vs EvoMAC (Hu et al., ICLR 2025b): EvoMAC relies on textual gradients generated by unit test feedback to evolve collaboration topologies, restricting its utility to tasks with deterministic code execution compilers; MA-Mesh establishes universal peer-reflection criteria applicable across multimodal tool use (GAIA) and scientific reasoning (GPQA).
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ Pioneering transfer of cloud-native service mesh and chaos engineering principles to LLM multi-agent orchestration.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive evaluation across four benchmarks, four synthetic chaos regimes, adversarial attacks, noisy tool settings, and real Kubernetes pod faults.
- Writing Quality: ⭐⭐⭐⭐⭐ Rigorous architectural formulation, crisp distinction between control and data planes, and transparent empirical ablation.
- Value: ⭐⭐⭐⭐⭐ Offers a foundational architectural blueprint for moving agentic systems from fragile local prototypes to high-availability distributed production environments.