Agent Trace RFC

i would like to have a way, at my day job, to glance at a PR diff and know whether certain lines were generated verbatim by claude. if i see a human being doing something that doesn’t match our best practices, i can imagine there are reasons they might have done so, and then try to understand them. perhaps they had a good reason that i don’t immediately see. if i know a line was fabricated without any underlying cognition, i can short circuit and just say “we don’t do things that way, do it this way and teach claude to do it this way in the future.”

maybe they will come back and tell me “no actually claude’s idea makes sense,” and at that point i can engage further. most of the time, it ends up with someone saying “oh that was just what claude came up with.” i find it frustrating when i waste time trying to understand why someone made a decision, when they didn’t actually make any decision at all.

…all that being said, i am not immediately convinced that the strategy outlined in this RFC is the right one to achieve those aims. but i do see value in line or even span level attribution for LLM generated text.

yes - somewhat. There are legal implications for one (potentially - depending on jurisdiction and the current perspective on what authorship means in the age of technology), as well as process perspectives (i.e. you can only improve that which you measure).