Knowledge-Based Zero-Replay Debugging of Multi-Agent LLM Traces.

arXiv:2606.14805v2 Announce Type: replace-cross Abstract: Reliable operation of multi-agent large language model (LLM) systems depends on debugging long execution traces, where the few causally decisive events are buried in unstructured logs of messages, route