Justin wrote the number on the whiteboard they didn't have. Instead, he typed it into a shared doc at 8:47 AM on Monday: 1,000. Below it, in smaller text: New scenarios. This week. Go.
Jay read the message from his desk and did the math. A thousand scenarios in five working days meant two hundred a day. They had 347 scenarios in the entire repository as of Friday's count. Justin was asking them to roughly quadruple the scenario corpus in a single sprint.
"He's lost it," Jay said, though he was already smiling.
Navan, who had been quietly cataloguing behavioral edge cases in the Slack twin since seven AM, looked up from his screen. "Has he, though? We're not writing code. We're writing descriptions of user intent. Natural language. The agents generate the implementation."
"You think we can actually do a thousand?"
"I think the agents can do a thousand. The question is whether we can describe a thousand distinct user journeys clearly enough for the agents to implement them correctly."
They started. Monday was slow. They managed sixty-three scenarios between the three of them, each one a carefully worded end-to-end user story. A new hire whose Okta provisioning triggers a cascade across six services. A manager who revokes access to a Google Drive folder and expects the downstream Jira permissions to update. An engineer who changes their display name in Slack and expects it to propagate everywhere.
By Tuesday afternoon they hit two hundred and the rhythm was locked in. Justin would describe a persona and an intent. Navan would identify the service boundaries the scenario would cross. Jay would formalize the satisfaction criteria—what observable outcomes would indicate the scenario was satisfied?
Then the agents took over.
On Wednesday morning, Jay noticed something odd in the logs. The agents had begun generating scenarios on their own. Not random ones—structured, purposeful scenarios that filled gaps in the coverage map. Scenario 412 tested a race condition between Okta deprovisioning and Jira ticket reassignment. Scenario 438 covered a Google Sheets permissions edge case that none of them had considered. Scenario 471 described a user who existed in Slack but had never been provisioned in Okta at all—a ghost user, an orphan identity.
"They're extrapolating from our patterns," Navan said, scrolling through the auto-generated scenarios with an expression that hovered between wonder and alarm. "They learned the shape of what we're testing and they're filling in the gaps."
By Wednesday evening the count was at 614, and only 283 of those had been written by humans.
Justin called a meeting. The three of them sat in the room with the long table, the monitor showing the scenario count climbing in real time. Every few seconds, the number ticked up. The agents were still going.
"Do we stop them?" Jay asked.
"Why would we?" Justin said.
"Because we said a thousand. They're going to hit a thousand by tomorrow morning. Without us."
Justin leaned back in his chair and laced his fingers behind his head. "The challenge was never about the number. The challenge was about whether we could teach the agents what a good scenario looks like." He nodded at the screen. "Looks like we did."
The count hit 1,000 at 2:17 AM Thursday. Nobody was awake to see it. By the time they arrived Friday morning, it was at 1,847, and the satisfaction metrics across the new scenarios were holding steady at 0.91.
Justin updated the shared doc. Below the original 1,000, he added a single line: Noted. Revising upward.
The moment where the agents start generating their own scenarios and filling coverage gaps is genuinely chilling in the best way. That's the real thesis of the factory, isn't it? You teach the shape, and the agents fill the space.