# 斯坦福与东北大学打造"AI 智能体版 Git"

- 来源：Rohan Paul (@rohanpaul_ai)
- 发布时间：2026-08-10 17:43
- AIHOT 分数：45
- AIHOT 链接：https://aihot.virxact.com/items/cmsn1riv504zyrogn47ks9fpb
- 原文链接：https://x.com/rohanpaul_ai/status/2086750257300529474

## AI 摘要

斯坦福和东北大学推出 Shepherd，一个“面向 AI 智能体的 Git”框架。它可将智能体的运行进程与文件系统一同提交，使执行过程可像 Git 一样回滚到任意历史提交并恢复。在双智能体协作测试中，Shepherd 通过监督回滚将性能差距缩小 91%（从 28.8% 提升至接近 57.2% 的单智能体水平）。

## 正文

Stanford and Northeastern built "Git for AI agents."

Solves a common problem with existing agent frameworks where record only transcripts and environment snapshots, so rewinding recovers files but not the live process or provider cache.

• It commits an agent's running process and filesystem together

• makes an agent's execution revertible like Git. Check out any past commit and resume from that exact state.

• turns every model action, tool call, and environment change into a typed commit.

• 2 coding agents splitting one repo normally score worse than a single agent doing both jobs alone: 28.8% versus 57.2%. Shepherd closes 91% of that gap by letting a supervisor watch both runs live and rewind one before a bad write lands, which is the clearest proof that reversible traces buy real performance, not just cleaner debugging.