Cihangir Bozdogan

What stands out: all 16 libraries converged on the same disaggregated architecture, but diverged sharply on staleness management and the hybrid (depth bounding + optional IS correction) trend feels right. Per-sample model_version tagging is the pragmatic foundation; once you have it, every other staleness strategy becomes a policy choice rather than an architectural rewrite. The MoE training-inference mismatch is the sleeper insight "Keep Routing" and "Keep Sampling Mask" stop being optimizations and become correctness requirements. Excellent survey.

Love the design choice of a single look_and_answer tool and letting Gemma 4 decide when vision is actually needed — much cleaner than always-on vision encoders. And going native llama.cpp over Docker is the right call if you want to actually swap the mmproj. A 5B Q4_K_M doing tool-routed multimodal reasoning on 8GB unified memory, with Parakeet + Kokoro on-device, is a strong signal for where edge agents are heading. Nice work.