GLM-5.3 begins optimizing its own system; Zhipu's Tang Jie: The minimal closed loop for RSI has emerged

Beating AI News Flash: Zhipu disclosed that the Infra Agent powered by GLM-5.3 has participated in optimizing the inference system of GLM-5.3-Flash. From first successful run to handling all production traffic in less than two weeks, end-to-end throughput increased to 3.2 times the initial level.

Co-founder and Chief Scientist Tang Jie noted in a retrospective that the place where Agents most easily get stuck is often not that they cannot write code, but that they do not know where the problem lies. For example, if a modification causes throughput to drop by 20%, it knows it broke something, but does not know which layer has the problem or what to check next.

So Zhipu turned the troubleshooting methods that senior engineers usually use into a set of tools that the Agent can directly use. After modifying code, the Agent can check on its own whether the results are correct, where time is being spent, and which approach is faster, then continue modifying based on the results.

With this method, the Agent identified issues such as long-context computation errors, KV transfer blocking, and repeated computation in decoding kernels. After one kernel was rewritten, it achieved a 1.71x speedup.

Tang Jie judged that the engineer's role will increasingly resemble "the person who designs feedback." Humans are responsible for setting goals, defining boundaries, and reviewing high-risk modifications, while the Agent proposes hypotheses, modifies code, and runs experiments on its own. It is still far from full RSI, but the minimal closed loop of "the model optimizes the system, and the system then serves the model" has already appeared.