OpenTelemetry tracing plus deliberate fault injection makes LLM multi-agent runs debuggable and A/B comparable

arXiv 2608.24271·medium signalXBlueskyLinkedInCopy link

llmmas-otel (arXiv 2608.24271, 2026-08-25) is a lightweight, framework-agnostic tool that instruments workflow phases, agent steps, inter-agent messages, tool calls and LLM invocations with OpenTelemetry distributed tracing, and pairs that with targeted fault injection. The point is reproducible comparison: run a baseline and a deliberately faulted execution, then inspect the difference through aligned traces and run artifacts. It was validated on both a minimal system and a real-world multi-agent development system, and it maps cleanly onto the standing rule that you should never trust a guard you have not watched fail.