vLLM

At KubeCon, NVIDIA donated OpenShell, an open-source runtime for securing and governing autonomous agents, to the CNCF as a Sandbox project, and joined Google's Agent Substrate framework for running long-lived agent workloads on Kubernetes. The move pushes agent governance and scale-out scheduling into open governance, alongside benchmarks showing Snapshot cut vLLM cold-start from minutes to tens of seconds.

Flide is a high-performance exact inference engine for Jev-style decision models, claiming up to 9.5x speedup over vLLM and SGLang via exact state forking while preserving single-question-equivalent outputs. It supports 30+ local decision models, exposes a /v1/systemone API, and installs via pip.