tl;dr: Local LLMs have crossed a usability threshold, with GPT-OSS and Google's Gemma models now enabling agentic coding workflows on a 64GB M2 Mac at roughly 75% the accuracy of frontier models. The author runs Gemma-4-12b-qat via LM Studio with the Pi agent harness in a Docker sandbox, successfully using it for refactoring, unit tests, and bootstrapping repos. Limitations remain—slow inference, small context windows, and occasional prompt template issues—but the tooling ecosystem (LM Studio, llama.cpp, HuggingFace) has matured enough to make local agentic development genuinely viable.