IQuest Research, in collaboration with institutions such as Beihang University and The University of Manchester, has introduced the ModularRSI framework. This framework allows agents to continuously iterate and optimize their peripheral operational mechanism, Harness, based on their own execution experiences while keeping the base model fully frozen throughout the process, thereby achieving recursive self-improvement. The ModularRSI framework subdivides Harness into five independent modules for separate evolution, identifies systemic flaws through multi-round trajectory comparisons, and integrates the capabilities of each module after screening specialized modifications through three rigorous validation mechanisms. In Terminal-Bench 2.0 testing, the framework successfully increased accuracy from 47.57% to 52.43%. Experimental results demonstrate that improvements at the Harness level possess cross-task and cross-domain transferability, even adapting to different base models that did not participate in the evolution. The study also found that task data of moderate difficulty is most conducive for Harness to extract general execution mechanisms, which evolve into reusable general execution rules rather than simple experiential memories, opening up a new path for recursive self-improvement of agents that does not rely on model weight updates.