[feature request] Specify different model type for each of subagents.

With the new subagent feature, it has become ready to assign some specialized subagents for analysis and planning, and forward the plan to other specialized subagents for execution.

If we could assume Sonnet to be able to perform better if given forwarded planning inputs generated from Opus, then we might got a chance to use Opus for the main agent and analysis and planning subagents, and use Sonnet for the execution subagents, and hopefully to match the same level of performance similar to all-Opus subagents.

Since each subagent has its own context, there is no concern about loss of KV-cache when switching to a different model, as a spawned subagent always has a fresh context.

I would assume this would be a very feasible idea, at least worth to explore with. If it can somewhat work, it could greatly reduce the burden on Anthropic's servers and also greatly saving API costs of end users.