How do you use subagents & multiple agent with local models, and how many?

Running qwen3.8 27b nvfp4 on vllm at max context only gives around 8 agents with 32k context each. That doesnt seem like much; what use cases do people use multi-agent frameworks and find it helpful for?