How do you debug something that is allowed to be wrong?

In a recent article on Dev.to, Pierre-Laurent Médori explores the complexities of debugging autonomous agents and LLM-based systems where errors are an inherent possibility. Rather than immediately blaming the model, Médori argues for a systematic approach to debugging that prioritizes verifying the underlying infrastructure. He introduces a framework based on three core questions: what the system is allowed to get wrong, which layer caused the error, and how long the error persisted. By maintaining an incident registry and implementing deterministic verification checks, developers can move beyond trusting transcripts—which often obscure the truth—and instead focus on measuring actual system behavior. The author emphasizes that debugging non-deterministic systems requires freezing verification tools and rigorously auditing data, configuration, and cache layers before questioning the prompt itself. Ultimately, the goal is not to eliminate all errors, but to significantly shorten the time to detection for failures.