📄 中文摘要
在构建AI代理时，开发者常面临“代理悖论”：代理在80%的情况下表现出色，但在20%的情况下却会出现严重错误。在生产应用中，80%的可靠性被视为失败。为了解决这一问题，开发者开始采用多代理编排和保护措施，而不是依赖一个庞大的“神代理”来处理所有任务。新的方法包括：路由器（快速小型模型，负责识别用户请求意图并将其发送给合适的专家）、工作者（针对特定任务进行微调的模型，如SQL生成或代码重构）和评论者（负责评估工作者的输出）。
自主工作流程的时代（为何80%的可靠性是失败）
出处: The Era of Agentic Workflows (and why 80% reliability is a failure)
发布: 2026年2月28日
📄 English Summary
The Era of Agentic Workflows (and why 80% reliability is a failure)
The construction of AI agents often encounters the 'Agent Paradox', where agents perform impressively 80% of the time but fail catastrophically 20% of the time. In production applications, this 80% reliability is considered a failure. To address this issue, developers are shifting towards multi-agent orchestration and guardrails instead of relying on a single 'God Agent' to manage all tasks. This new approach includes: The Router, a small and fast model that determines the intent of user requests and directs them to the appropriate specialist; The Worker, a model fine-tuned for specific tasks such as SQL generation or code refactoring; and The Critic, which evaluates the outputs of the Worker.