# Google 过程图如何改进长程智能体

- 来源：DAIR.AI (@dair_ai)
- 发布时间：2026-09-10 02:37
- AIHOT 分数：44
- AIHOT 链接：https://aihot.news/items/cmtugppcx1cherofptc5k6k0g
- 原文链接：https://x.com/dair_ai/status/2097756208857837851

## AI 摘要

Google 提出 Procedural Graph，将长程智能体的过程知识显式化为 procedure-relation-procedure 三元组，供智能体查询下一步动作及条件。框架通过 LLM 精炼器对比失败与成功轨迹并自我改写图结构，仅保留通过验证的编辑。从最小骨架出发，可构建媲美或超越手工设计的图，并修复有缺陷的专家先验。

## 正文

Recommended paper from Google if you want to learn how to improve long-horizon agents with procedural graphs

https://x.com/omarsar0/status/2097755424007373270?s=20

### 引用推文

> elvis：Another banger paper from Google. If you build memory for long-horizon agents, this one is worth your time. (bookmark it) Really nice to see how knowledge graph...