Software Factory Archive

Jay made a graph. This was not unusual. Jay made graphs the way other people made coffee—reflexively, habitually, as a way of processing the world through the lens of structured data. But this graph was different, because this graph had a milestone on it, and the milestone was one million dollars.

The factory had crossed one million dollars in cumulative token spend. One million dollars fed into language models that had, in return, produced tens of thousands of lines of working code, hundreds of validated scenarios, and a satisfaction metric that had climbed steadily from 0.4 in July to 0.91 in January.

Jay's graph had two lines. The blue line was cumulative token spend. It rose in a smooth, steady incline—roughly a thousand dollars per day per engineer, three engineers, plus overhead for batch runs and experimental pipelines. The red line was satisfaction, inverted and normalized to the same scale. As spend went up, the cost per unit of satisfaction improvement went down. The lines crossed in October, which was the point at which the factory became, by Jay's calculation, more cost-efficient than a traditional engineering team of equivalent output.

"One million dollars," Jay said, presenting the graph at the morning not-a-stand-up. "That's a lot of tokens."

"What did we get for it?" Navan asked, not because he didn't know but because he wanted Jay to say it out loud. He liked the way Jay articulated these things. There was a precision to Jay's summaries that turned data into narrative.

"We got a factory that produces validated software at a rate no human team can match, with a defect rate lower than any published benchmark for AI-assisted development, at a declining marginal cost per unit of improvement." Jay pointed at the graph. "The cost per satisfaction point has been declining steadily since August. We're spending more in absolute terms, but we're getting exponentially more value per dollar."

"Justin's rule," Navan said. "A thousand dollars per day per engineer."

"At this point, we should be spending more," Justin said. He said it without looking up from his screen. "The return on additional token spend hasn't plateaued. The curve is still improving. We should be pushing until we find the inflection point where additional spend stops producing proportional improvement."

"You want to spend more money," Jay said.

"I want to find the ceiling. We haven't found the ceiling yet."

Jay added a new line to his graph: projected spend at increased rates. The projection showed the cost curve continuing to decline for at least another quarter before flattening. They were nowhere near the ceiling.

Navan took a photo of the graph and taped it into his notebook next to a handwritten note: "One million dollars in tokens. Zero dollars in regret."

Justin looked at Jay's graph one more time. "Make it a dashboard," he said. "Real-time. I want to see the cost-per-satisfaction curve every morning."

Jay made it a dashboard. The agents wrote the code. It cost twelve dollars in tokens and took fourteen minutes. The irony was not lost on anyone.