This hits a ache level I have run into individually. I added LangSmith tracing to my very own agentic undertaking in particular as a result of when one thing broke mid-session I had no solution to reconstruct which instrument name or retrieval trail brought about it.. simply vibes and scrolling. Re_gent seems like that very same observability thought however implemented on the report/code degree, which is in reality the place it hurts extra. The prompt-to-diff linking is the characteristic I would use maximum.. realizing which instruction brought about a particular report alternate is if truth be told helpful for debugging agent loops that move improper on step 4 of seven.
Curious how you are dealing with branching despite the fact that.. if an agent takes two other approaches throughout periods at the identical report, does re_gent observe the ones as separate branches or does it flatten the whole lot into one linear historical past? That difference issues so much as soon as brokers get started doing exploratory paintings.
