What I've done so far is to set up 3 classes of subagents: main (kimi), chat (qwen) and deep (Claude). The main agent is supposed to delegate nearly every basic request down to qwen .. which is really cheap. My startup, config and debug have all been with Kimi, which is the one I interact with. I scuffled a bit, driving the cost up. Since it got stable it's been more reasonable. But not yet coding anything. Mostly research and productivity tasks. One thing my agent struggles with is commitment to delegation. If the first results from qwen aren't exactly what I wanted, the agent tries to just do it themselves.. instead of improving the prompt and rerunning. Trying to break it of that habit.