Agent Observability Without Logging Chain-of-Thought

When agents fail, engineers often ask, "What was it thinking?" This confuses two distinct needs: researchers studying model reasoning and operators needing to understand the execution context. Logging hidden chain-of-thought doesn't reliably answer operational questions. Instead, agent observability should focus on execution explanations. This approach emphasizes three types of records: traces for operations, events for state changes, and audit receipts for accountability, enabling better diagnosis and understanding of agent performance.