Harness-Aware Self-Evolving Improves AI Model Performance

Traditional self-evolving AI frameworks typically focus on optimizing task solutions while treating the surrounding "harness" – the environment, tools, or prompts guiding the AI – as a fixed element. This research introduces Harness-Aware Self-Evolving (HASE), a novel agentic reinforcement learning framework that allows a single model to simultaneously generate task solutions and dynamically edit selected components of its own harness. HASE operates within a multi-turn action space, enabling the model to adapt both its internal workings and its external operational context. The effectiveness of HASE was demonstrated across several challenging tasks. For instance, a Qwen3-8B model utilizing HASE achieved text-classification performance comparable to a much larger GPT-OSS-120B model that relied on a separate Claude Code harness proposer. Furthermore, HASE significantly outperformed reported baselines in alpha factor mining and successfully repaired imperfect evaluation components, converging to state-of-the-art performance in circle-packing algorithm discovery. These results underscore HASE's ability to improve both the AI's solution and its operational harness through a unified, self-optimizing agentic process.