📄 中文摘要
大型语言模型（LLM）展现出卓越的能力，但其固有的概率性特性给企业级部署带来了严峻的可靠性挑战。为解决此问题，六西格玛智能体被提出，这是一种新颖的架构，通过三个协同组件实现企业级可靠性。首先，任务被分解为由原子操作组成的依赖树，确保复杂任务的结构化处理。其次，采用微代理采样机制，每个任务并行执行n次，并利用多样化的模型、提示词和推理策略，以最大化结果的多样性和鲁棒性。最后，通过共识驱动的验证和聚合过程，对多个并行执行结果进行评估、交叉验证和合并，从而生成高度可靠的最终输出。这种方法显著降低了LLM固有的不确定性，使其能够满足企业应用对高准确性、一致性和可信赖性的严苛要求。
📄 English Summary
The Six Sigma Agent: Achieving Enterprise-Grade Reliability in LLM Systems Through Consensus-Driven Decomposed Execution
Large Language Models (LLMs) demonstrate remarkable capabilities, yet their fundamentally probabilistic nature poses critical reliability challenges for enterprise deployment. To address this, the Six Sigma Agent is introduced, a novel architecture designed to achieve enterprise-grade reliability through three synergistic components. First, tasks are meticulously decomposed into a dependency tree of atomic actions, ensuring structured processing of complex operations. Second, a micro-agent sampling mechanism is employed, where each task is executed 'n' times in parallel, leveraging diverse models, prompts, and reasoning strategies to maximize result variety and robustness. Finally, a consensus-driven validation and aggregation process evaluates, cross-validates, and merges the multiple parallel execution outcomes to generate highly reliable final outputs. This methodology significantly mitigates the inherent uncertainties of LLMs, enabling them to meet the stringent demands for high accuracy, consistency, and trustworthiness in enterprise applications. By integrating task decomposition, diversified execution, and consensus aggregation, the Six Sigma Agent provides a robust solution for building LLM systems that can operate stably in critical business scenarios. The proposed architecture aims to facilitate the widespread adoption of LLMs in reliability-critical domains such as finance, healthcare, and industrial automation, thereby overcoming major obstacles in current LLM deployment.