papersSEP 10 04:00 UTC
RobustSGPO adds search-space control to semantic-gradient prompt optimization for agent harnesses
A new arXiv preprint presents RobustSGPO, an extension of semantic-gradient prompt optimization for evolving agent harnesses with execution feedback. The method addresses an unresolved part of SGPO's local update rule—deciding which portion of the prompt to edit and what kind of change to make—by constraining the search space of candidate edits. According to the authors, specifying the requested edit at each step makes automated harness improvement more controlled.