Software Factory Archive

Jay discovered the pattern at 3:47 AM on a Tuesday, which was itself evidence for the pattern.

He'd been up late—not working, exactly, but watching. The dashboard had become a habit. He'd check it before bed, and sometimes the checking became watching, and sometimes the watching lasted until the sky started to lighten. This particular night, he was tracking a cluster of stubborn scenarios that had been oscillating around 0.78 for days, refusing to converge.

At 3:12 AM, all six of them jumped. Not incrementally. They leapt—0.78 to 0.86, 0.77 to 0.89, 0.79 to 0.91. Simultaneously. As if some constraint had been lifted, some interference pattern had dissolved.

Jay pulled the logs. The agent runs during the 3-5 AM window were qualitatively different from the daytime runs. Faster convergence. Cleaner code generation. Fewer oscillations in the feedback loops. The same scenarios that flailed during business hours resolved smoothly in the small hours of the morning.

He went back through three months of data. The pattern was consistent. The best convergences—the highest-quality agent outputs, the steepest satisfaction improvements—clustered between 3 AM and 5 AM Pacific time.

Jay did what any engineer would do. He wrote a paper about it.

The paper was eight pages. It had charts. It had a methodology section. He titled it "Temporal Patterns in Autonomous Agent Convergence Rates: Evidence for Off-Peak Computational Advantages" and it contained, among other things, a compelling statistical analysis showing that agent performance correlated inversely with Anthropic API load, which he hypothesized peaked during US business hours and troughed in the pre-dawn window.

"It's not mystical," he told Navan the next morning, sliding a printout across the table. "It's infrastructure. Lower API latency at 3 AM. Faster round-trip times. The agents get more iterations per unit time, and each iteration is faster, so the convergence loops tighten."

Navan read the paper. "This is good," he said. "This is publishable."

Justin read the paper at lunch. He was quiet for a while, which Jay had learned to interpret as either deep appreciation or deep concern. It was deep concern.

"Delete it," Justin said.

"What?"

"If you publish this, two things happen. One: every team using Claude for agentic workflows shifts their heavy compute jobs to the 3-5 AM window. Two: the 3-5 AM window stops being off-peak. You've discovered an arbitrage, Jay. The moment you publish the arbitrage, it disappears."

Jay stared at him. "You want me to suppress a research finding for competitive advantage."

"I want you to not set our own infrastructure on fire." Justin slid the printout back. "Also, you misspelled 'computational' in the abstract."

Jay deleted the paper. He did not, however, delete the dashboard view that tracked convergence rates by time of day. And he did not stop checking the dashboard at 3 AM. If anything, he checked it more often now, with the guilty pleasure of someone who knows a secret about the universe that they're not allowed to tell anyone.

The 3 AM window held. Every night, while the team slept, the agents found their best selves in the quiet hours. Jay watched them do it, alone, from his apartment, the dashboard glowing in the dark like a fire nobody else could see.