The NeoHorse-1 project introduces a framework for recursive self-improvement (RSI) via agentic post-training, utilizing a heterogeneous model pool managed by an intelligent routing harness. The system captures user interaction data—including capability demand, service tier selection, and reasoning traces—to construct high-quality training examples that integrate interleaved tool calls and harness context. These examples undergo a rigorous admission process involving structural validation, subscene-level labeling, and six-dimensional semantic evaluation. The training methodology follows a three-stage curriculum driven by routing signals, which extends to routing-guided on-policy distillation where teacher models supervise student responses within the established progression. To close the learning loop, capability-guided allocation utilizes evaluation feedback to dynamically adjust the training mixture for subsequent iterations. Empirical results across eleven benchmarks, including coding, tool use, and instruction following, demonstrate significant performance gains. Specifically, post-training increased the macro-average scores from 58.94 to 64.87 for the 4B model and from 65.60 to 69.04 for the 9B model. Notably, this process effectively narrows the performance gap between the post-trained 4B model and the 9B base model, providing a viable prototype for harness-mediated, feedback-driven model evolution.