SGLang Summit: agents now write ~95% of merged PRs as v1.0 and Miles v0.2 ship

hsu_byron · x · 2026-11-13

SGLang v1.0 and the Miles RL framework v0.2 shipped together at the SGLang Summit.

Agents as primary contributors: official slides show 95% of merged PRs in SGLang are now agent-written, versus all hand-written a year ago. The surrounding automation includes 36 skills in the repo, 328 PRs reviewed via the SGLang Review Inbox, 200K lines refactored by agents, and a 2.1x gain from the Muse Spark open-source megakernel.
SGLang v1.0: high-performance serving for LLMs and multimodal models, with zero-overhead scheduler, full-graph coverage, speculative decoding, unified memory, PD disaggregation, and hierarchical cache.
Miles v0.2: RL for LLMs and diffusion across NVIDIA/AMD/TPU/Trainium, four modalities, and Multi-LoRA. Periodic Labs uses it for long-horizon scientific RL, cutting checkpoint conversion from 30 minutes to 1 and upstreaming the change.

The poster's takeaway: whatever the noise, frontier AI infra keeps moving.

More from coding & agent