Philipp Schmid(@_philschmid)

In today's harness engineering experiment, i ran an eval comparing two bash execution styles, persis...

6.4内容质量

TL;DR · AI 摘要

Philipp Schmid on X: "In today's harness engineering experiment, i ran an eval comparing two bash execution styles, pers...

核心要点

  • 主题聚焦：In today's harness engineering experiment, i ran
  • 来源：Philipp Schmid(@_philschmid)，建议结合原文判断细节。
  • AI 分析暂不可用，本条为保底评分与摘要。
#AI
打开原文

Philipp Schmid on X: "In today's harness engineering experiment, i ran an eval comparing two bash execution styles, persistent (cd and environment variables survive) and clean stateless. Persistent mode sounds obvious at first. You change directories once and keep working. But this can break the" / X

Philipp Schmid

@_philschmid

In today's harness engineering experiment, i ran an eval comparing two bash execution styles, persistent (cd and environment variables survive) and clean stateless. Persistent mode sounds obvious at first. You change directories once and keep working. But this can break the agent's spatial awareness. Persistent worked fine when every action stayed inside bash. They broke when the agent used other tools with “paths” tools like Read or Write. After cd app/data in bash it tried to call Write(path="settings.conf"). I also tested whether stripping - working directory: <path> from SI would change it. It fixed the spatial confusion of cd + read, but worsen any other eval as the agent has no idea in where currently is.

6:03 PM · Jul 15, 2026

4.4K

Views

9

2

8

28

Read 9 replies