Agent-Based Development Approach for SGLang

SGLang development goes beyond simple code modification, encompassing complex processes spanning LLM serving, distributed runtimes, GPU kernels, Diffusion pipelines, and production incident response. In the past, these workflows relied on developers' personal experience and memory, but now they are being converted into executable SKILL.md files and scripts for agents to utilize.

The current SGLang ecosystem has built the following skills for agents.

• .claude/skills: Manages repository-level development workflows including CUDA crash debugging, kernel integration, CI, and profiling
• Diffusion-specific skills: Adding new Diffusion models, benchmarking denoise paths, verifying quantization pipelines
• KDA-Pilot: Applies KDA-style agent kernel workflows to SGLang to track kernel work and performance improvements

The true value of agents comes from Procedural Engineering Knowledge that includes executable steps, reproducible experiments, and reviewable evidence. In particular, performance optimization work is evolving into a Loop Engineering stage where benchmarking, profiling, patching, and revalidation are repeated.

As a result, the developer's role is shifting away from simple implementation and toward Review-centered work: defining problems for experiments performed by agents, selecting evidence, designing workflows, and determining whether outcomes are production-ready.