Microsoft Research, in collaboration with Shanghai Jiao Tong University, Tongji University, and Fudan University, has introduced SkillOpt—a text-space optimizer designed to enhance AI agent performance without retraining the underlying model.
SkillOpt trains a single natural-language skill document while keeping the target model frozen. An optimizer model reviews scored rollouts and proposes bounded edits—additions, deletions, or replacements—to the skill text. A held-out selection split accepts an edit only if it strictly improves the score. The final output is a single artifact: best_skill.md.
Transfer Evaluation Methodology
The transfer experiments report three key metrics: Baseline, the target model’s score without any skill; Direct, where SkillOpt is trained in-domain on the exact target model; and Transferred, where a skill trained on a different model is applied without further optimization. The critical comparison is not simply transferred versus direct, but rather how much of the in-domain gain survives the transfer.
Cross-Model Transfer: Mixed Retention Within a Family
In tests involving GPT-5.4 and its smaller variants, skills optimized on the larger model were deployed on smaller ones. Results on SpreadsheetBench showed GPT-5.4-mini achieving a baseline of 36.1%—but full details of retention rates are provided in the original paper.
Cross-Harness Transfer: From Codex to Claude Code
A significant finding is that skills optimized for OpenAI's Codex harness can be transferred to Anthropic's Claude Code harness with minimal performance loss. This suggests that skill artifacts are not tightly coupled to the execution environment, opening the door for cross-platform skill reuse.
Looking Ahead: 2026 Context and Implications
As AI agents become more complex and widely deployed, the ability to reuse optimized skills across models and harnesses is increasingly valuable. SkillOpt addresses a key bottleneck: the cost of per-model optimization. By demonstrating effective transfer, it hints at a future where skill libraries are portable assets, potentially decoupling skill development from specific model architectures or vendor tools.
For developers and enterprises, this means reduced overhead when upgrading models or switching between coding assistants. The approach also aligns with broader industry trends toward modular, reusable AI components—an area of growing importance in 2026 as organizations seek to maximize ROI on AI investments.
While SkillOpt’s current results are promising, the team acknowledges that transferability varies across tasks and model families. Future work will likely explore broader benchmarks and more diverse harnesses to fully map the boundaries of skill transferability.
For the full research details, including comprehensive tables and experimental setups, refer to the original paper via the provided link.
via MarkTechPost