GLM-5.3 agent optimizes own inference system 3.2x

Zhipu AI used a GLM-5.3-driven Infra Agent to optimize its own inference system on domestic accelerators. Within two weeks, end-to-end throughput increased 3.2x. The agent found three real issues: TF32 precision drift, GIL blocking, and redundant normalization, demonstrating a closed loop of model optimizing system.