@myx I work with this daily. Totally agree. The gap is the context (not LLM context) that missing. If I debug logs in the same session I wrote the code, I get a pretty decent outcome. I think that specifying scope and context to an LLM will heavily improve it. The big win of agents vs humans is that the LLM can parse a huge block of logs and work with it well. Humans will read line 1 and generally make an assumption and jump into debug - LLM can be told to go parse the full log set and find patterns first, identify false positives, and identify pre-existing unrelated noise.