跳转至

OrchestraBench:评估多智能体编排的失效模式、恢复能力与任务拆解质量

文章背景与核心概要

随着多智能体编排框架从探索性演示(Demo)加速迈向生产环境,现有的基准测试通常仅衡量最终任务的准确率。这导致它们往往无法诊断流水线在 何处 崩溃、系统级联故障是从哪里开始的,或者到底是哪项路由决策导致了故障。

为了填补这一空白,OrchestraBench 引入了一种基于模板化企业工作流设计的、受控且支持种子复现的故障注入测试框架。该基准测试通过级联半径(cascade radius)按失效模式分类的恢复率(per-failure-mode recovery)等核心指标,并辅以自助法置信区间(bootstrap confidence intervals)与配对统计检验,全面评估故障传播、恢复率以及任务拆解质量。


📌 Summary / 摘要

As multi-agent orchestration frameworks transition from exploratory demos to production environments, current benchmarks typically measure only end-task accuracy. They often fail to diagnose why a pipeline broke, where systemic cascades began, or which routing decisions caused the failure.

To bridge this gap, OrchestraBench introduces a controlled, seed-reproducible failure-injection harness designed over templated enterprise workflows. The benchmark evaluates failure propagation, recovery rates, and task decomposition quality using metrics such as cascade radius and per-failure-mode recovery, backed by bootstrap confidence intervals and paired statistical tests.

随着多智能体编排框架从探索性演示(Demo)加速迈向生产环境,现有的基准测试通常仅衡量最终任务的准确率。这导致它们往往无法诊断流水线在 何处 崩溃、系统级联故障是从哪里开始的,或者到底是哪项路由决策导致了故障。

为了填补这一空白,OrchestraBench 引入了一种基于模板化企业工作流设计的、受控且支持种子复现的故障注入测试框架。该基准测试通过级联半径按失效模式分类的恢复率等指标,并辅以自助法置信区间与配对统计检验,全面评估故障传播、恢复率以及任务拆解质量。


🔍 Key Findings & Results / 关键发现与结果

  • Routing Policies: On a 26-case gold-labelled diagnostic, a naive keyword/flag router scored 0% on adversarial test cases featuring missing or misleading surface flags. Conversely, an intent-reasoning model router achieved a 100% success rate, matching the performance of an oracle.
  • Failure-Handling Tiers: Controlled mechanism probes utilizing a real Claude agent across a verifiable arithmetic dependency chain revealed three distinct failure-handling tiers across five MAST modes:
  • Tool faults: Recovered fully (1.0)
  • Ambiguous delegation: Recovered partially (0.30)
  • Latent or semantic modes: Never recovered (0.0)
  • Generalizability: This performance hierarchy remained consistent when the computational task was reframed into a loan-approval workflow, and across different model scales (Claude Sonnet, Opus, and Haiku), although absolute recovery rates shifted depending on the context window.
  • Impact of Blind Retries: Blind retry strategies successfully reproduced latent faults and increased the time-to-detection, highlighting that active detection and attribution are prerequisites for containment.
  • Cascade Dynamics: The cascade radius scaled upward with pipeline depth, expanding from a mean of 0.9 to 4.7 across depths ranging from 3 to 7.
  • State Repair Ablation: A trusted-state repair ablation demonstrated that apparent containment improvements stemmed primarily from the trusted-state signal itself rather than autonomous system detection.

(Note: These findings are derived from controlled-chain mechanism probes rather than broad domain-workload claims.)

  • 路由策略: 在包含 26 个黄金标准标签的诊断测试中,针对存在表面标志缺失或误导的对抗性测试用例,朴素的关键词/标志路由器(keyword/flag router)得分为 0%。相反,意图推理模型路由器(intent-reasoning model router)实现了 100% 的成功率,性能与理想预言机(oracle)不相上下。
  • 故障处理层级: 利用真实的 Claude 智能体在可验证的算术依赖链上进行的受控机制探测,揭示了在五种 MAST 模式下存在三个截然不同的故障处理层级:
  • 工具故障(Tool faults): 完全恢复(1.0
  • 模糊委派(Ambiguous delegation): 部分恢复(0.30
  • 潜在或语义模式(Latent or semantic modes): 从未恢复(0.0
  • 泛化能力: 当计算任务被重构为贷款审批工作流,且跨越不同的模型规模(Claude Sonnet、Opus 和 Haiku)时,这种性能等级层级依然保持一致,尽管绝对恢复率会根据上下文窗口的大小而发生变化。
  • 盲目重试的影响: 盲目重试策略成功重现了潜在故障并增加了检测时间(time-to-detection),这凸显了主动检测与归因是实现故障控制(containment)的前提条件。
  • 级联动力学: 级联半径随流水线深度的增加而扩大,在深度从 3 到 7 的范围内,平均值从 0.9 扩展到 4.7
  • 状态修复消融实验: 受信任状态修复的消融实验表明,表面上的故障控制改善主要源于受信任状态信号本身,而不是系统的自主检测能力。

(注:以上发现均源自受控链条机制探测,而非对广泛领域工作负载的泛化断言。)