Justin drew three curves on the whiteboard. He used three different colored markers. Blue, red, green. He labeled them with the careful handwriting of someone who'd been a CTO long enough to know that illegible diagrams waste everyone's time.
The blue curve was compound interest. Slow start, then exponential. Everyone knew this curve. It was on the cover of every personal finance book ever written.
The red curve was compounding error. Same shape, same exponential growth, but pointing toward disaster. Every bug that introduced another bug. Every quick fix that created two new problems. Every shortcut that accumulated into technical debt until the codebase was a landfill with a REST API on top.
"This is where most software projects live," Justin said, tapping the red curve. "They start clean. They accumulate small errors. Each error makes the next error more likely. The curve steepens. Eventually the project is more debt than product."
Jay and Navan had seen this before. Every engineer had. It was the central tragedy of the profession: the thing you built started dying the moment you stopped paying perfect attention, and nobody could pay perfect attention.
"The third curve," Justin said, moving to the green line, "is compounding correctness."
He drew it. Same exponential shape. But this time each correct decision made the next correct decision more likely. Each clean module made the next module cleaner. Each passing scenario made the satisfaction metric more reliable, which made the next round of agent work more constrained, which made the output better.
"Before October 2024, this curve was theoretical," Justin said. "Human developers couldn't sustain it. We'd hit the red curve eventually, no matter how disciplined we were, because humans get tired. Humans have deadlines. Humans take shortcuts."
"And the agents don't," Navan said. Not a question.
"The agents don't get tired. The agents don't have Friday-afternoon energy. The agents don't merge code because the sprint ends tomorrow. They run the scenarios. They compute satisfaction. They iterate until they converge. And every correct iteration feeds the next one."
Jay studied the three curves. Blue: money compounds. Red: error compounds. Green: correctness compounds. Three exponentials. The same math driving all of them. The only difference was what you were compounding.
"The factory lives on the green curve," Justin said. "That's the whole thesis. Long-horizon agentic coding workflows that compound correctness rather than error. Everything else is implementation detail."
He capped the markers and set them in the tray. The whiteboard stayed up for weeks. People walked past it and stopped. Some took photos. The three curves became a kind of shorthand—a team Rorschach test. When someone talked about the old way, they'd gesture at the red line. When someone talked about the factory, they'd point to the green.
Navan eventually redrew it in his notebook, in pencil, with better proportions. He added a fourth curve, very faintly: the x-axis. The flatline. The project that never compounds anything at all.
He didn't show anyone that one either.
Navan's secret fourth curve (the flatline) is the most devastating thing in the archive. Some projects really do compound nothing. They just exist, flat, forever. I have worked on that project.