Zhipu AI GLM-5.3-Flash Inference System Self-Optimized by AI Agent

Zhipu AI announced that its GLM-5.3-Flash inference system, running on over 100,000 domestic AI accelerators, was optimized by an AI agent driven by GLM-5.3 itself. The system went from first run to full production in under two weeks, with end-to-end throughput improved 3.2x. The agent resolved three real engineering issues including precision drift and concurrency bottlenecks.
Key insights
Agent-Driven Optimization Achieves 3.2x Throughput Gain
Zhipu AI disclosed that the GLM-5.3-Flash inference system, deployed on over 100,000 domestic AI accelerators, was optimized by an AI agent powered by GLM-5.3. From initial run to full production took less than two weeks, achieving a 3.2x end-to-end throughput improvement. The agent performed optimizations such as ReplaySSM (compute for memory), intra-node tensor parallelism (communication for memory), mixed INT8/FP8/BF16 caching (precision for capacity), and encode-prefill-decode separation (separation for scheduling flexibility).
Dense Feedback System Enables Agent to Debug Complex Issues
The agent's key challenge was not writing code but diagnosing why performance degraded. Zhipu AI built a layered verification interface providing dense feedback: correctness (is the computation correct?), system behavior (where is time spent?), and performance (which solution wins under what conditions). Each signal is local, cheap, and objectively verifiable. This allowed the agent to identify and fix three real problems: TF32 rounding error accumulation in KDA kernels, GIL blocking in KV transfer, and redundant normalization in decode kernels.
Three Case Studies Demonstrate Agent Capabilities
The agent discovered and fixed three distinct engineering issues. First, precision drift in KDA context-parallel paths due to TF32 rounding error accumulation in chained state matrix merges. Second, KV transfer and DeepEP dispatch could not overlap because the intra-node path did not release the GIL. Third, a decode kernel recomputed the same normalization four times due to blocking. Each case represents a different capability: numerical correctness attribution, cross-language system debugging, and pattern transfer across codebases.
Engineer Role Shifts from Problem Solver to Feedback Designer
Zhipu AI CEO Tang Jie emphasized that the boundary remains: humans define goals, build feedback environments, and review high-risk changes. The engineer's role is evolving from solving problems to designing feedback systems. The layered verifiable feedback environment serves dual purpose: it is both the agent's working environment and training material for next-generation models. Each completed task becomes training data for successors.
Recursive Self-Improvement Loop Exists but Still Distant
Tang Jie stated that the smallest closed loop already exists: the model optimizes the system, and the system runs the model. However, recursive self-improvement remains far off. The dense feedback environment is designed to be reused for training future models, creating a data flywheel. The approach is seen as a step toward autonomous AI development.