Meta and University of Illinois researchers build EvoHarness-RL for self-evolving agent runtime

Meta and researchers from the University of Illinois have jointly developed EvoHarness-RL, a framework designed to let large language model agents create, access, and manage external state systems on their own. According to Techub, citing CryptoBriefing, the framework lifted success rates on the ALFWorld benchmark from 47.9% to 96.9%, a sharp jump in task completion performance. The system builds a structured external state through three components: Belief, Progress, and Experience. It is trained with supervised fine-tuning and GRPO optimization. The report said this setup gives agents a way to organize and use information outside the model itself while handling multi-step tasks. Researchers also observed two spontaneous behaviors during training, described as "tool annealing" and "tool evolution." These behaviors allowed agents to gradually internalize operations and refine the structure of their external state. The work centers on runtime design for AI agents and how those agents can improve the way they use tools and state over the course of training.