Microsoft's Skillopt Enhances AI Agent Skills via Open Source

WHO:

VentureBeat - Yifan Yang, Senior Research SDE at Microsoft Research Asia, is involved in the development of SkillOpt.

WHAT:

VentureBeat - SkillOpt is a new open-source framework developed by Microsoft that optimizes AI agent skills without changing model weights.

VentureBeat - SkillOpt significantly boosts accuracy for models like GPT-5.5 and Qwen, outperforming existing baselines.

VentureBeat - SkillOpt was tested across 52 evaluated combinations of model, benchmark, and harness, proving highly effective.

VentureBeat - SkillOpt delivered an average absolute improvement of +23.5 points against the no-skill baseline on GPT-5.5.

VentureBeat - GPT-5.4-nano nearly doubled its score on multimodal document QA and tripled its score on embodied interaction and sequential decision-making.

VentureBeat - The final deployed skills never exceeded 2,000 tokens, with a median length of roughly 920 tokens.

WHERE:

VentureBeat - SkillOpt is designed for use in various execution environments, including Codex CLI and Claude Code.

WHEN:

VentureBeat - Open-source developers are scheduling SkillOpt to run periodically over their agents' past trajectories.

WHY:

VentureBeat - The framework aims to improve the reliability of AI agents in automating complex enterprise workflows, such as document data extraction and compliance tasks.

VentureBeat - The optimization process addresses the challenges of manually updating agent skills, which is often slow and error-prone.

HOW:

VentureBeat - SkillOpt uses deep-learning-style optimization to systematically explore modifications to skill documents based on performance feedback.

VentureBeat - The optimization process involves an iterative propose-and-test loop that separates the model executing tasks from the model optimizing the skill.

VentureBeat - SkillOpt filters proposed edits to avoid duplicates or contradictions and ranks them by expected utility before applying changes.

VentureBeat - The framework integrates smoothly with existing orchestration stacks, allowing for harmonious operation with other systems like DSPy.

(UPDATED June 11, 2026 19:55 UTC)

Microsoft has introduced SkillOpt, an innovative open-source framework designed to enhance AI agent skills without altering model weights, and Yifan Yang, a Senior Research SDE at Microsoft Research Asia, is leading its development. This framework has shown remarkable results, delivering an average absolute improvement of +23.5 points on GPT-5.5 and significantly boosting performance across various tasks, including nearly doubling scores on multimodal document QA. SkillOpt is set to be utilized in diverse execution environments and aims to streamline the optimization of AI agents, making them more reliable in automating complex enterprise workflows.