Software Factory Archive

Jay noticed it first.

He was reviewing the morning's agent output—not the code itself, because that was against the rules, but the metrics. The satisfaction curves. The convergence graphs that told the story of what the agents had done overnight while the three of them slept in their respective apartments scattered across the city.

"Justin," he said, not looking up from his monitor. "Come look at this."

Justin was at his standing desk, annotating something on his tablet. He walked over, coffee in hand, and leaned over Jay's shoulder. The graph on screen showed scenario pass rates over the past six weeks. It was a standard chart, one they checked every morning. But today it looked different.

"That's not linear," Justin said quietly.

"No," Jay agreed. "It's not."

The curve was bending upward. Not dramatically—not like a hockey stick in a VC pitch deck. It was subtle, the kind of thing you'd miss if you weren't staring at these numbers every single day. But it was unmistakable. The rate of scenario satisfaction wasn't just increasing. The rate of increase was increasing.

Navan rolled his chair over from the other side of the room. "What are we looking at?"

"Compounding," Justin said. He took a sip of coffee. His expression was the one Jay had learned to read over the past two months—the one that meant Justin was running calculations behind his eyes that he wouldn't share for another forty-eight hours.

Jay pulled up the diff logs. Not the code—the structural summaries the agents produced alongside their commits. He scrolled through the last week. The agents had refactored the Jira twin's webhook handler three times. Each refactor was smaller than the last. Each one resolved a class of edge cases rather than individual failures.

"They're not just fixing bugs," Jay said slowly. "They're learning which kinds of fixes prevent future bugs."

"That's what compounding correctness means," Justin said. "I've been waiting to see if it would actually happen."

Navan was quiet for a moment. He pulled up his own dashboard—the DTU health metrics—and flipped to the Google Sheets twin. Same curve. He switched to Okta. Same curve, offset by about four days.

"It's happening across all the twins," Navan said. "Not just the scenario layer."

The three of them sat there, the hum of the server rack filling the silence. Jay thought about the October 2024 revision of Claude 3.5, the moment Justin always pointed to as the beginning. The moment long-horizon agentic coding workflows began to compound correctness rather than error. Justin had used that exact phrase in the founding document. Jay had thought it was aspirational. Marketing language dressed up as engineering philosophy.

He didn't think that anymore.

"So what do we do?" Navan asked.

Justin set his coffee down. "We don't touch it. We don't interfere. We watch, and we make sure the scenarios keep getting harder." He paused. "If the agents are compounding, the worst thing we can do is flatten the curve by making their job easy."

Jay looked back at the graph. The curve bent upward, patient and relentless, like compound interest on a debt the universe owed to correctness.

He saved a screenshot. He had a feeling they'd want to remember this morning.