Tag
The author investigates how small the orchestration model in a local-first agent can be while maintaining trustworthiness, finding that tool-call discipline breaks before reasoning, and a 35B-A3B MoE model proves viable on 12GB hardware.