Tencent's Hyra-1.0: An AI Agent That Gets Better on Its Own

Tencent's Hunyuan team launched Hyra-1.0, the first version of an AI agent that can improve itself recursively. It's built for research and engineering tasks that need high performance. Hyra works well in standard tests, but it also learns in real-world AI development, science, and industry settings.

The design follows the 'Bitter Lesson' – keep the framework simple and give the agent a wide range of actions. Given a task, Hyra runs a loop: it explores and creates better solutions based on past experience. This continues until the agent stops or runs out of budget. The best solution from the whole process is returned. For example, Hyra can take a 2D reference image, design a 3D model, and generate a renderable version.

Recursive self-improvement (RSI) and automated research are hot topics in AI. Google DeepMind's AlphaEvolve, Together AI's collaborative agents, and Andrej Karpathy's autoresearch all show how agents can now beat humans in open scientific discovery.