RobustSGPO improves AI agent prompt editing and task success rates
RobustSGPO: Search-Space Control for Agent Harness Evolution
Artificial Intelligence
Summary
Improving AI agents often involves tweaking how they are instructed or prompted. The authors found that a common method can get stuck by only making local, small changes without clear rules about how to edit. They created RobustSGPO, a method that controls where and how edits happen, checks these edits before applying them, and manages different starting points during the search. This approach led to better performance on a set of brainstorming tasks, increasing how often tasks were completed and improving overall quality.
What this means in practice
- •For ai tool developers: Enhance AI agent behavior by applying controlled, executable prompt edits that improve task completion rates and output quality.
- •For workflow automation engineers: Improve the reliability of AI brainstorming and task-solving workflows by using permission scheduling and progressive search-space control.
Authors
Zibo Zhao, Jijun Shi, Mo Zhou, Zhongyuan Wang, Shifu Bie, Yunfei Zhang, Xuanting Zhou, Xiangyu Wu, Bin Liu, Ruiming Tang, Wenwu Ou, Kun Gai
Abstract
Semantic-gradient-based prompt optimization (SGPO) improves agent harnesses using execution feedback, but its local update rule leaves the choice of edit scope and operation unresolved. We introduce RobustSGPO, which specifies the requested edit, constructs and checks the patch, and continues search from either the incumbent or retained snapshots. We evaluate permission scheduling, cumulative controls, and task-family transfer in the AgentX brainstorming workflow using 120 tasks, 95 runs, and 7,350 candidate attempts. Periodic $1\to2\to3$ scheduling exceeds fixed maximum permission by 0.28 test-score points. RobustSGPO increases completion on 30 held-out tasks from 60.0% to 80.0% and improves test quality from 3.77 to 4.14 under a 20-million-token budget. Category retention reduces source-task degradation after a shift, whereas random retention reaches a higher destination endpoint. Search-space control benefits quality through executable edits and alternative starting points, with measurable retention overhead.