My AI agent orchestrator had a conflict of interest: same model family was writing code and reviewing. Models from the same family share assumptions and blind spots, so I handed the review to a different vendor (Cursor instead of Claude) to de-risk the family bias.