A user reports sustained performance of ~68 tokens per second running Qwen3.8 Flash Next (4.7bpw MLX quantization) as a local agent on a MacBook Pro M5 Max with 128GB RAM. The setup uses mlx-serve and the Hermes Agent framework with tool calling and Obsidian integration for long-context automation tasks (100k-190k tokens). The author notes mlx-serve outperforms LM Studio for agent workloads and shares insights on making local agent automations reliable.