Why can I not replay exactly what my agent did when it failed in production?

An LLM agent that fails in production cannot be reliably reproduced for debugging because re-running the same prompt yields different outputs. Most agent frameworks log inputs and final outputs but discard the full execution trace, meaning every intermediate tool call, model response, and state mutation is gone the moment the run ends. Without execution lineage, a failed run is a forensic black hole and engineers reconstruct failures from symptoms rather than causes. A May 2026 paper proposes deterministic execution graphs as a first primitive for this but notes that capturing and replaying non-deterministic traces at production scale without prohibitive overhead is still an open engineering problem. A separate 2026 case study of agentic failures in scientific workflows found that plausible-looking but wrong intermediate steps propagated silently through pipelines precisely because no re