I can't keep track of which AI models come from where and which ones are suited for what. I'm in the camp that wants model selection handled well by auto.
Writing out the exact input/output types and failing test cases before prompting is usually the only way to keep an agent from inventing its own architecture.