we've been building local agents with vision models, works great for gui automation but coding tasks still need cloud models for reliability