User: local sub-agent extensions ignore VRAM limits

A user reports that sub-agent implementations for local LLMs do not account for VRAM constraints, unlike cloud-based use with GPT-5.4 and Sonnet. The insight contrasts work and home setups on limited hardware.