Verification is the hard half of agents

There is a strange asymmetry in how teams build with agents today. Generation, getting a model to produce plausible actions, is close to free. Verification, knowing whether those actions were correct, safe, and repeatable, is where all the real cost lives.

The demo cliff

Every team we talk to has an impressive demo. The same team, three weeks later, has a support channel full of "the agent did something weird and we don't know why". The demo proves generation works. Production is a verification problem.

What verification actually needs

It needs durable runs so you can inspect what happened. It needs traces so you can see every step. It needs guardrails so the risky actions pause. And it needs evals in CI so a regression is caught before a customer feels it. None of that is the model's job.

That is the half Nebula owns.