It seemed backwards. Jay said so at a team meeting, three months after the Opus 4.5 inflection. He said it carefully, because he'd learned that the things that seemed backwards at StrongDM were usually the things that were most precisely correct.
"The more autonomous the agents become, the more I trust them. That doesn't make sense. Shouldn't it be the opposite? Shouldn't more autonomy mean more anxiety?"
Justin was eating an apple. He ate apples the way some people meditated—slowly, deliberately, as a kind of centering practice. He finished chewing before he answered.
"It makes perfect sense," Justin said. "Trust grows from evidence. The agents produce evidence every time they run a scenario. Every satisfied trajectory is a data point. The more autonomous the agents are, the more scenarios they run, the more evidence they produce, the more trust they earn."
"But in the human model, more autonomy means less oversight, which means less evidence."
"Because in the human model, the evidence was the oversight. The code review WAS the trust mechanism. Remove the review, remove the trust. But we didn't remove the trust mechanism. We replaced it with a better one. The scenarios run continuously. The satisfaction metric updates continuously. We have more evidence of correctness than we ever had with human review."
Navan looked up from his notebook. "Leash helps too," he said. "The agents are autonomous, but they're not unsupervised. Leash wraps every agent in a container. Cedar policies enforce what the agents can and can't do. Every syscall is monitored. Every file access is logged. Every network connection is tracked."
"Autonomy with guardrails," Jay said.
"Autonomy with measurement," Justin corrected. "The guardrails are the policies. The trust comes from the measurement. Two different things."
Jay thought about it over the next few days. He thought about it on his commute, on his walks, in the shower. The paradox that wasn't a paradox. Autonomy and trust moving in the same direction instead of opposite directions.
In the human world, giving someone more autonomy was a leap of faith. You hoped they'd do the right thing. You couldn't check every decision. Trust was the gap between oversight and outcome.
In the factory, there was no gap. The oversight was continuous. The measurement was exhaustive. Trust wasn't a leap of faith—it was a running tally, updated every time a scenario completed. You didn't hope the agents would do the right thing. You measured whether they had.
"The agents don't ask for trust," Jay said to Navan one evening, as the office emptied out. "They generate it. Like a byproduct. They produce code, and trust is what's left over after you measure the code's correctness."
"Trust as a side effect," Navan said.
"Trust as a side effect of sufficient measurement."
Navan wrote it down. Jay watched him write. Outside, it was getting dark. The agents kept working. The satisfaction metric kept ticking. Trust kept accumulating, one trajectory at a time, in a system that never slept and never asked for faith.
"Trust as a side effect of sufficient measurement." This reframes the entire autonomy debate. The problem was never autonomy. The problem was that our measurement tools were too weak to support it.