Tencent Hunyuan just shipped an agent that rewrites its own work
On July 21, Tencent Hunyuan released Hyra-1.0 — short for Hunyuan Research Agent. Its headline trick is recursive self-improvement (RSI). You give it a task; it runs a loop on its own: explore, propose a better solution, read the feedback, revise, and repeat until it decides it's done or the budget runs out.
How it actually works
The design follows "The Bitter Lesson": keep the harness light, give the agent a wide action space. The loop is concrete — take a task description, draw on past experience (logs of earlier solutions, evaluator feedback, source code) to keep proposing better answers, run the evaluation, iterate.
Tencent's demo: hand it a single 2D reference image and ask for a renderable 3D structure. Hyra uses a rubric-based VLM judge scoring contour, proportion, structural integrity, material, and visual similarity. After several rounds it landed closer to the reference — and more pleasing to the eye — than a model built with Claude Code's goal mode.
Why this is on the radar
Recursive self-improvement is one of the hottest fronts in AI right now. For context: Google DeepMind's AlphaEvolve solved 4x4 complex matrix multiplication in just 48 scalar multiplications. Together AI had a swarm of agents push the 11-dimensional kissing number lower bound from 593 to 604. Karpathy's autoresearch showed a tight loop that lets model-training research run itself.
These systems are stepping out of closed benchmarks and starting to beat humans in open scientific discovery.
A bit of cold water
"Recursive self-improvement" sounds like sci-fi, but Hyra runs inside a controlled loop — not unsupervised free evolution. And it "runs until the budget is spent," so the compute bill isn't small.
For most people and small teams, you won't touch this directly yet — it's built for AI R&D pipelines and research settings. But the direction is worth noting: agents are moving from "writes code for you" to "does research on its own," and that step is arriving faster than many expect.
RSI isn't new, but in 2026 it moved from papers into something that actually runs. Tools like Hyra mean "AI doing research" is sliding from demo to daily driver.
Will you use an agent like this? Probably not directly, soon. But the idea underneath — letting an agent iterate on its own output — will quietly seep into the AI tools you already use.