Interesting idea, but worth flagging the security trade-off before this becomes a default pattern: every manual “ok, go ahead” is also a checkpoint. Right now, a human eyeballing the output between steps is often the only thing standing between a compromised or hallucinated intermediate result and the AI acting on it downstream.
I agree with you. Human approval at each important step is still an important safety checkpoint, especially when AI can make mistakes or produce compromised results.
Completely agree, and I think this is such an important point to keep raising as automation increases. Human approval at each meaningful step isn’t a bottleneck to eliminate it’s genuinely one of the most valuable safety nets we have. It gives us a real chance to catch a hallucinated result or a compromised intermediate output before it cascades downstream into something harder to undo.
Thanks for raising this. Human review is one of the safeguards covered in our developer guidance: https://developers.openai.com/api/docs/guides/safety-best-practices. We appreciate the discussion about where checkpoints are most useful.