# RobustSGPO：为 Agent Harness 进化补上搜索空间控制

- 来源：DAIR.AI (@dair_ai)
- 发布时间：2026-09-10 23:15
- AIHOT 分数：41
- AIHOT 链接：https://aihot.news/items/cmtvonjl40glcronbpa6apbv4
- 原文链接：https://x.com/dair_ai/status/2098067735389593926

## AI 摘要

RobustSGPO 为语义梯度提示优化补上了编辑控制机制：明确请求的编辑内容、在接受前构建并检查补丁，并从当前方案或保留快照继续搜索。在 AgentX 头脑风暴工作流中，120 个任务、95 次运行、7350 次候选尝试的测试显示，30 个留出任务的完成率从 60.0% 升至 80.0%，测试质量从 3.77 升至 4.14（2000 万 token 预算）。

## 正文

If you run automated prompt or harness evolution, this one is worth your time.

(bookmark it)

Semantic-gradient prompt optimization improves an agent harness from execution feedback, but its local update rule never decides how large an edit to request or which operation to apply.

RobustSGPO adds that.

It specifies the requested edit, constructs and checks the patch before accepting it, and continues search from either the incumbent or a retained snapshot.

Measured over 120 tasks, 95 runs and 7,350 candidate attempts in the AgentX brainstorming workflow, completion on 30 held-out tasks rises from 60.0% to 80.0% and test quality from 3.77 to 4.14 under a 20-million-token budget.

Periodic 1 to 2 to 3 edit-permission scheduling beats fixed maximum permission by 0.28 test-score points, so how much the optimizer is allowed to change per step is itself worth scheduling.

Retention is a real trade. Category retention reduces source-task degradation after a task-family shift, while random retention reaches a higher destination endpoint, and both carry measurable overhead.

Chat with Paper: https://academy.dair.ai/papers/robustsgpo-search-space-control-for-agent-harness-evolution-2609.09646