Agent Loops: Transforming Tokens into Productivity

​📣 Agent Loops: Transforming Tokens into Productivity

​SGLang Office Hours is back!

​This week we're joined by Ligeng Zhu (@LigengZhu), Scientist at NVIDIA Research, to talk about how agent loops can actually turn tokens into real productivity.

​Ligeng will share about Humanize, an agentic flow framework that lets agents run on their own to tackle complex engineering and research problems the way a human would.

​He'll then walk through three long-running agent-loop case studies:

• ​KDA: writing fast CUDA kernels, ranked #1~3 on the MLSys FlashInfer Kernel Contest

• ​Anthropic's Virtual Hardware: optimizing computation on hardware that doesn't physically exist yet

• ​JetAutoResearch: ahead-of-time compilation that cuts >50% of the AutoResearch workflow cost

​He'll also talk about what's coming in Humanize 2.0 / 3.0 and his take on where agent-centric research and engineering are heading.

​Join our office hour stream and bring your questions for Ligeng!
Youtube Living stream 👉 SGLang Office Hour


Join SGLang Slack 👉 http://slack.sglang.ai/
Following us on X 👉 https://x.com/lmsysorg

​If this helps you, please consider giving us a star — it truly motivates the team. ⭐ https://github.com/sgl-project/sglang