On July 21, Tencent Hunyuan officially released Hyra-1.0 (Hunyuan Research Agent) — the first domestic Recursive Self-Improvement (RSI) agent targeting research and engineering tasks. Hyra's design follows The Bitter Lesson: keep external frameworks light, make the action space wide. Given a task description, it runs an exploration loop continuously — using historical experience (including the run logs of past solutions, evaluator feedback, and source code) to propose better solutions until it actively ends or the budget runs out, finally returning the historically best solution. This "experience reuse + automatic reflection" pattern lets Hyra evolve continuously not only on public benchmarks but also in product systems, AI R&D pipelines, and even natural science and industrial scenarios. The official demo shows the advantage of this capability: given only a 2D reference image to generate a renderable 3D model, Hyra iterates across multiple rounds modifying modeling code and rendering parameters, scored by a rubrics-based VLM judge on silhouette, proportion, structural integrity, material and other dimensions. The result is closer to the reference image and more human-aesthetic than Claude Code's goal mode. Over the past year, RSI and automated research have been mostly led overseas: DeepMind's AlphaEvolve did 4×4 complex matrix multiplication in 48 scalar multiplications; Together AI pushed the 11-dimensional kissing number lower bound to 604; Karpathy's autoresearch proposed a streamlined training self-loop paradigm. Tencent's entry this time means domestic vendors are starting to systematically fill in this frontier — and the real test of RSI is "running long, steady and safe in production". Hyra hangs verification directly onto R&D pipelines and industrial scenarios, an attitude itself more worth watching than a pure demo.