🛠 SGLang Shifts to Agentic Development
The SGLang team is implementing a "skills" system (SKILLS.md), where engineering knowledge is encoded into executable procedures for AI agents. Claude Code and Codex are being used to automate profiling processes and CUDA crash debugging.
🌍 This approach transforms high-performance AI software development from a "craft" into a scalable process, allowing for faster code optimization for new hardware, such as the NVIDIA B200.
👤 LLM tool development will become faster: new optimizations and attention types will reach production frameworks much more rapidly thanks to automated testing cycles.
Source 1: https://www.lmsys.org/blog/2026-07-02-agent-assisted-sglang-development/