بچه‌های بخش تحقیقات گوگل یه مقاله تازه دادن بیرون به اسم WikiSkill که قشنگ بازی ایجنت‌های هوش مصنوعی رو عوض می‌کنه

Google Research just published a fascinating paper on self-improving AI agents. "WikiSkill" lets agents turn their own experience into persistent knowledge that compounds over time. Execution traces are analyzed, useful patterns are stored in a wiki, and that knowledge is then used to continuously create and improve reusable agent skills. The results are impressive: Qwen 3.5 9B with evolved skills reached 47.4% average accuracy, beating Qwen 3.6 27B without skills at 39.4%. On some individual benchmarks the gains were huge, with Qwen 3.6 27B jumping from 40.8% to 81.7% on SpreadsheetBench. Even more interesting: skills transfer between models. In some cases, a model performed better using skills learned by another model than skills it developed itself. This feels like an important direction for agents: not just better models, but agents that accumulate experience, turn it into persistent knowledge, and actually get better at their jobs over time.