In the first hour, they defined the constraints. Justin was specific about this. Not goals, not aspirations, not a roadmap. Constraints. The things they would not do.
"Constraints are what make a system," he said. "Every interesting system in nature is defined by what it can't do, not what it can. A river is the shape of the rock it can't flow through."
Navan wrote that down. Jay suspected he would have written it down even if it hadn't been profound. The kid wrote everything down.
Constraint one: no human-written code. They had covered this. The whiteboard still bore its two commandments in block letters, the ink already drying into the faint permanence of marker that's been left uncapped too long.
Constraint two: no human code review. Also covered, but Justin elaborated now. "The moment a human reviews code, you've created a bottleneck that scales at human speed. The factory's throughput is bounded by the slowest reader in the room."
"So what validates correctness?" Jay asked. He was sitting with his arms folded, not defensive but structural, like he was bracing for something.
"Scenarios," Justin said. "End-to-end user stories. Not unit tests. Not integration tests. Scenarios that describe what the user experiences. The agents run the scenarios and we measure satisfaction—what fraction of observed trajectories satisfy the user's intent."
"Probabilistic," Navan said, looking up from his notebook.
"Yes. Not pass-fail. Probabilistic. Because the world is probabilistic and any validation framework that pretends otherwise is lying."
Constraint three: the scenarios live outside the codebase. "Like holdout sets in machine learning," Justin said. "The agents never see the test data during development. The scenarios are sacred ground."
Constraint four, and this was the one that made the air in the room change: spend at least one thousand dollars per day per engineer on tokens.
Jay unfolded his arms. "A thousand dollars a day."
"Minimum. If you're spending less than that, the factory has room for improvement. It means you're not pushing enough work through the agents. It means somewhere in the pipeline, a human is doing something a machine should be doing."
There was a long silence. Outside, a garbage truck backed up with a rhythmic beeping. The monitor on the wall displayed a terminal cursor, blinking patiently, like a metronome for a song no one had started playing.
"This is different," Navan said. He said it flatly, not with wonder or skepticism, but with the measured tone of someone naming an observable fact.
"This is different," Justin agreed.
Jay looked at the whiteboard, at the monitor, at his two new colleagues. He'd spent fifteen years writing code. Go, mostly. Careful, well-tested, documented Go. Libraries that other people depended on. He'd built things he was proud of with his hands on a keyboard.
Now he was being told that his hands didn't belong on the keyboard anymore. That the most valuable thing he could do was describe the world accurately enough that something else could do the building.
The air had changed. He could feel it. Something about the room was no longer hypothetical.
"Okay," Jay said. "Show me what a scenario looks like."
"A thousand dollars a day." The way Jay unfolds his arms at that line. That's the physical reaction of someone recalibrating fifteen years of professional instinct in real time.