Self-Improving Agents

Recursive self-improvement architectures emerge: AREX with inner/outer loops, SAO, Ornith, Active-GRPO. SkillRise introduces unified RL for cross-task skill evolution. NeoCognition $40M seed. New benchmark InMind exposes implicit-association blind spot in agent memory. Shadow evaluation of AI research agents shows they handle engineering but fail at creative research design and backtracking — highlighting current limitations for open-ended self-improvement.