Microsoft and Google Release Open-Source AI Tools for Optimizing Agent Performance

Microsoft Research Asia has open-sourced SkillOpt (MIT License), a framework that treats AI agent skill documents — typically markdown files containing procedural instructions — as trainable objects, optimizing them through an iterative propose-and-test loop without modifying underlying model weights. The system borrows concepts from deep learning, including an edit budget analogous to a learning rate, held-out validation gates, and a rejected-edit buffer acting as negative memory, to prevent the instability that plagues manual or loosely controlled skill updates. In benchmarks across 52 combinations of model, harness, and task, SkillOpt delivered an average absolute improvement of +23.5 points over a no-skill baseline on GPT-5.5, and skills trained in one execution environment transferred to others — a spreadsheet skill trained in Codex CLI yielded a +59.7 point gain when moved to Claude Code. Separately, Google DeepMind released DiffusionGemma, a 26-billion-parameter mixture-of-experts model that generates text by iteratively denoising a canvas of random tokens rather than producing them autoregressively, making it predominantly compute-bound rather than memory-bandwidth-bound. This architecture allows it to run on as little as 18 GB of DRAM or VRAM and achieve up to approximately 4x faster output than a comparable autoregressive model on a single Nvidia H100, though Google acknowledges it trails Gemma 4 12B on quality benchmarks such as GPQA-Diamond. DiffusionGemma is available on Hugging Face under an Apache 2.0 license with support in vLLM, MLX, and HF Transformers, and is positioned as an experimental rather than enterprise-grade release.