Microsoft has introduced SkillOpt, a new framework designed to enhance AI agent skills by treating the skill document as the primary focus of training, rather than altering the AI model itself. This approach allows for better task handling as SkillOpt monitors agent performance, learning from successes and failures, and applies a more powerful optimization model to refine skills based on established benchmarks. In testing, SkillOpt excelled across various scenarios, improving accuracy by 23.5 points for GPT-5.5, the latest model from OpenAI released in April 2026, which is tailored for complex tasks. This advancement plays into Microsoft’s broader AI strategy aimed at developing self-evolving agent capabilities, aligning with their recent expansion of Copilot Studio.

Codex: Codex refers to an agent execution setting used in evaluations of AI models for code-related tasks, associated with OpenAI’s coding-focused capabilities. It was one of the three agent settings tested in the SkillOpt research. The news indicates SkillOpt performed best or tied across cases including this setting.
GPT-5.5: GPT-5.5 is OpenAI’s frontier model released in April 2026, designed for complex professional tasks including coding and reasoning. It supports advanced agent-like workflows with strong performance on multifaceted problems. The news notes SkillOpt’s effectiveness when tested with GPT-5.5 in direct chat settings.
SkillOpt: SkillOpt is a text-space optimization system for training reusable skills in AI agents, introduced in a May 2026 arXiv paper by Microsoft-affiliated researchers. It treats skill documents as trainable artifacts edited by a separate optimizer model while keeping the base agent frozen. The approach is highlighted in the news for enabling reliable improvements across diverse benchmarks and models without model retraining.
Microsoft: Microsoft is a major technology company advancing AI research and agent technologies through its research division and products like Copilot. Recent efforts focus on developing practical tools for AI agents in enterprise and developer settings. In the news, Microsoft researchers introduced SkillOpt to improve how AI agents handle skills through external document optimization.
Claude Code: Claude Code denotes a code-oriented agent configuration leveraging Anthropic’s Claude models. It served as one of the agent settings in the SkillOpt paper’s experiments. SkillOpt demonstrated strong results across this and other settings according to the provided research.

Research Focus: Microsoft’s recent AI work emphasizes self-evolving capabilities for agents, aligning with the SkillOpt methodology for stable skill improvements.
Model Ecosystem: OpenAI released GPT-5.5 in April 2026 as a new frontier model optimized for complex professional and coding workloads.
AI Agent Platforms: Microsoft expanded Copilot Studio in May 2026 with generally available computer-using agents that interact directly with websites and desktop applications.