Agent S4's evolved interfaces reportedly boost a 27B model's OSWorld V2 score 3.8-fold

A post about Simular's Agent S4 says a meta-agent reads task trajectories without success labels and builds interfaces, with unseen validation tasks selecting which versions survive. It reports Qwen3.8-27B rising from 5.6% to 21.3% on OSWorld V2, above the post's stated GPT-5.5 and Claude Opus 4.8 scores, at $1.10 per task. The post also reports ScienceBoard success rising from 41.4% to 58.6%.