LiteLLM and MCP show where infrastructure and tool permissions escape their intended boundaries. This issue also covers language-sensitive guards, reproducible RAG, mobile inference, and four runtime updates.
This week: agents that fail on repeat, Pydantic AI security fixes, and SGLang's new storage default. Plus runtime snapshots, reusable context, and audio models running on a phone's CPU.
MCP security advisories, agent evaluation pitfalls, and runtime changes in vLLM, Ollama, and Apple's on-device stack. Plus a proposed MLX memory experiment.
