At the 2026 Physical AI and World Models conference in Nanjing, Agnes AI (爱思科技) livestreamed a 5v5 Honor of Kings match in which an AI coach rewrote the team's tactical playbook after each game, lifting its win rate against built-in bots from 0% to 100% over five revisions, according to the company's booth demo reported by 新智元. No model parameters were trained in that experiment; the coach, described as the forthcoming Agnes 3.0 Pro, only edited prompts, action choices and code, with each version required to pass researcher-set tests before play. A separate demo showed a Franka Panda arm in MuJoCo simulation updating its own weights during work via Agnes Embed TTT, raising task success from roughly 30% to about 90% while cutting iteration time to as little as 30 minutes to an hour. Agnes frames both routes as part of its open-source Agnes Harness agent infrastructure and its "万物共脑" (one brain for everything) pitch, and says a version of the harness also coordinated simulated Mars-base devices and booth robots.