Agentic Fine-tuning with RFT

Hi, have a bunch of multi-turn agent traces (approx 12-16 steps long) generated from o4-mini, want to try the new RFT fine-tuning API. In the examples posted online, the setup is single-turn convo, but since I am using for an agentic usecase, curious if submitting an entire agent trace of multiple Q/A’s along with a grader for the entire agent sequence would work well? Or am I better off isolating and picking out specific agentic turns?