顶级工程师经验：可靠的智能体系统依赖确定性步骤而非堆叠LLM
Saravia 和一堆顶级工程师踩坑后的共识：agent 系统里多放确定性步骤，错误才不会滚雪球。还在往工作流里疯狂塞 LLM 的人应该停下来看看，做分类器和验证器才是提升可靠性的正道。
Elvis Saravia（omarsar0）在 X 上分享了他与多位顶级工程师构建智能体系统的共同经验：表现最好的 agentic 系统包含大量无聊的确定性步骤，而不是用更多 LLM agent 覆盖每个环节。
其核心论点是可靠性问题：单纯堆叠 LLM agent 无法提升系统可靠性，因为每一步的错误会复利式放大成更多错误。要让智能体工作流朝正确方向复利，仍然需要搭配分类器、验证器和评估器等传统确定性组件。
这条推文获得 29 次点赞、22 次收藏，评论区 21 条讨论，反映出社区对“LLM 不是万能锤子”这一工程共识的共鸣。对正在搭建生产级智能体工作流的开发者而言，这条经验指向了混合架构——LLM 负责灵活推理，确定性组件负责质量控制。
This might surprise many new to AI. The best agentic systems I have built and seen other top engineers build involve lots of boring deterministic steps. You can't just throw more LLM agents at every problem and expect reliability to improve. Every error will compound into more errors. You still need good classifiers, verifiers, and evaluators for agentic workflows to truly compound in the right direction. This is key to your "software factory."📊 22 ⚡
This report is categorized under AITOP's "How-to" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include agents, prompt engineering, LLM, agentic workflow.
Source attribution: curated from elvis (@omarsar0).
