Recursive self-improvement and RL are becoming the next agent-training frontier

MiMo-V2.6, Sakana AI's MASS, Video-RSI, Recurrent Harness, Stateless Language Agents, and QwenGyre show rapid progress toward agents that improve through environments, evolving workflows, persistent state, and infrastructure-aware RL. Reported efficiency and benchmark gains are substantial, but evaluator gaming, benchmark regressions, synthetic-data dependence, reproducibility, and real-world transfer remain unresolved.