Engineers are sharing strategies for debugging AI workflows after users reported sudden cost spikes in production. The consensus emphasizes using detailed logging and tracing to isolate bottlenecks like infinite loops, excessive retries, or context bloating.