Browser LLMs go truly local. Local speech-to-speech matures. Local-first agent workflows multiply. Benchmarks and plumbing realities. Coding agents: speed, skills, and guardrails