Debugging agents is the practice of tracing and evaluating an agent's full execution path to find why it failed. Developers inspect model calls, prompts, retrieved context, tool inputs and outputs, state changes, retries, policies, and final responses.
Agent debugging requires observability. Without traces and spans, teams are left guessing whether the model failed, the tool failed, the context was missing, or the orchestration loop made the wrong decision.