China is leaning into recursive self-improvement! Tencent’s Terminal-RSI gives us a concrete look at how that works. The idea is simple. Let a Qwen-based agent tackle real terminal tasks. Take the runs that pass verification, rewrite them into reproducible demonstrations, and train the agent on them. Refine it with reinforcement learning. Then put the improved agent back to work generating examples for the next round. After two cycles, average success across four terminal benchmarks went from 20.68% → 50.58%. On Terminal-Bench 4.0, it went from zero to 19.84%. On the long-horizon benchmark, it went from zero to 5 of 46 tasks fully completed. That’s real progress, with most of the hardest work still ahead. The part worth watching is what happens on the next pass. Can a better agent discover better training examples often enough to keep the gains coming?