- Published on
Microsoft’s SkillOpt Shows Optimized Agent Skill Artifacts Transfer Across Model Scales and Between Codex and Claude ...
- Authors
- Name
- DeepHorus
Microsoft researchers, in collaboration with Shanghai Jiao Tong University, Tongji University, and Fudan University, have introduced SkillOpt, a text-space optimizer that makes AI agent skills transferable across model scales and execution harnesses. The method trains a single natural-language skill document while keeping the target model frozen, allowing a portable artifact to be reused with minimal performance loss. This development could significantly reduce the cost of optimizing AI agents for new tasks.
SkillOpt works through an iterative editing process. An optimizer model reads scored rollouts and proposes bounded add, delete, or replace edits to the skill document. Only edits that strictly improve performance on a held-out selection split are accepted. The final artifact is a single markdown file, best_skill.md, which can be consumed directly by various AI agent harnesses without special formatting.
Cross-model transfer tests show that skills optimized on a large model like GPT-5.4 can be effectively deployed on smaller variants. For example, SpreadsheetBench on GPT-5.4-mini retains 82% of the in-domain gain, nearly matching the performance of direct optimization. In a surprising case, LiveMath on GPT-5.4-nano achieved a transferred score of 28.8, exceeding the in-domain SkillOpt result of 27.2, suggesting some learned procedures are model-agnostic.
The most significant finding is cross-harness transfer. A skill optimized inside Codex lifted Claude Code's SpreadsheetBench score from 22.1 to 81.8, slightly surpassing the 80.4 achieved by training a skill directly in Claude Code. While the two harnesses expose different tool APIs and command surfaces, procedural skills like structure-first inspection and formula-aware verification remain portable because they operate on the workbook level, independent of the CLI.
However, not all skills transfer equally. LiveMath retained only 10% of the in-domain gain when moving from Codex to Claude Code, indicating that reasoning-heavy skills are more tied to their training environment. The paper also notes that cross-family transfer, such as from GPT to Qwen, was not tested, and retention is not uniform across benchmarks or model sizes.
Cross-benchmark transfer produced positive but modest gains. On Omni-MATH, improvements over the no-skill baseline were observed at all three model scales, even though no in-domain SkillOpt run was reported. The authors attribute this to reusable mathematical procedures that survive changes in test instances and answer-format conventions.
The portability of SkillOpt artifacts stems from a shared contract: all execution modes—direct chat, Codex, and Claude Code—consume the same best_skill.md format. Final skills are compact, ranging from 379 to 1,995 tokens with a median near 920, and consist of 1 to 4 accepted edits. Crucially, the learned rules are procedural rather than instance-specific, as shown by the SpreadsheetBench rule: inspect workbook structure and formulas, then write evaluated static values across the full target range—a strategy that works regardless of the harness.
These findings suggest that procedural skills represent the portable class of agent knowledge, while reasoning-heavy skills require more targeted optimization. As AI agents are increasingly deployed across diverse platforms, transferable skill artifacts like those from SkillOpt could enable efficient cross-platform reuse and reduce the need for task-specific fine-tuning. Future research may explore cross-family transfer and additional benchmarks to further validate this approach.
