9B small model autonomously updates intelligent agent skills, with profits equaling Claude's flagship large model According to Beating monitoring, large model intelligent agents often undergo self evolution by updating external "equipment" (Harness, including prompt words, memory, skills, and tools). The latest paper "Harness Updating Is Not Harness Benefit: Disrupting Evolution Capabilities in Self Evolving LLM Agents" released by institutions such as Pennsylvania State University, AmericaC, and Amazon first decouples the evolution process into two dimensions: "equipment updates" of the evolutionist and "equipment benefits" of the executing end. Cross testing shows that the equipment update capability exhibits a clear "flattening" feature on the model base capability. The maximum profit gap brought by updating equipment with different models is only 3.1%, and even the updated skills produced by Qwen3.5-9B with a size of 9B are highly equivalent in program structure to the flagship Claude Opus 4.6. This indicates that when developing self evolving systems, there is no need to spend high cost models on the roles of the evolutionists. On the contrary, the ability of intelligent agents to benefit from equipment shows a non monotonic trend. The performance of top-level models is approaching the ceiling, while weak models (such as Qwen3-32B) have the greatest potential for improvement but actually benefit the least. The study points out that weak models have two major failure modes. Firstly, there is the issue of 'equipment activation failure'. The weak model has a skill loading rate of only 25.1% in the SkillsBench benchmark, while the strong model has a loading rate of around 96%. Next is' equipment compliance failure ', as the long-range execution trajectory unfolds, the instruction compliance of weak models will sharply drop from 0.52 in the initial loading stage to 0.13. The cutting-edge discovery has resonated strongly with renowned AI researcher Elvis Sar (@ omarsar0). Elvis Sar pointed out that he observed a completely consistent phenomenon in his experiments on encoding intelligent agents and long-range tasks: more powerful models do not always evolve into better intelligent agents. This paper provides important guidance for the architecture design of intelligent agent systems, indicating that the computing budget should be tilted towards executing agents, and the focus should be on strengthening the equipment's ability to autonomously wake up and follow long-range instructions during agent training. [Original link]