Open-Ended Optimization: Rethinking Self-Evolving Agents Without Prescribed Pipelines

A recent study published on arXiv (2608.09629v1) questions the traditional framework for self-evolving agents, which often depend on fixed optimization processes that govern evidence collection, artifact modification, candidate selection, and criteria for completion. The authors explore whether such specific procedures are essential when a frontier model serves as the optimizer. They propose Open-Ended Optimization (OEO), a strategy that maintains the objective, interaction limits, resource allocation, data constraints, and evaluation methods, while enabling the optimizer to develop the enhancement process dynamically. OEO was tested against two established methods: SkillOpt, a staged pipeline with limited edits, and GEPA, a reflective evolutionary approach. In 14 direct comparisons across 8 benchmark-target-model scenarios, GPT-5.5-driven OEO secured 12 victories, 1 tie, and 1 close defeat by 0.21 percentage points, utilizing a median of 34.3 percent of SkillOpt's target-interaction token budget. A one-shot, zero-interaction control is also referenced, although specifics are truncated. The findings indicate that adaptable, open-ended optimization may surpass rigid pipelines while being more efficient in resource use.