NeoHorse-1 — harness-mediated agentic post-training and self-improvement

Agent improvement increasingly combines harness-policy optimization, synthetic environments, data allocation, and literature-driven changes. Self-HPO reports coding resolve rates rising from 57.8% to 64.8%; CoworkEnv reports a 17.5-point Claw-Eval SFT gain; ScholarEvolve reports Qwen3.5-27B gains on AppWorld and Tau2-Bench without weight changes, while a safeguard example shows harness improvements can regress legitimate tool use. ToM2Vec offers a related, distinct lead in latent partner modeling; replication, human-partner tests, transfer, costs, and causal attribution remain open.