The data was unambiguous. Jay had charted it across six weeks, binning agent output by hour, and the pattern was always the same: the agents were most productive between midnight and five AM Pacific time. Not by a small margin. By forty percent.
More scenarios completed. Higher satisfaction scores. Cleaner code. Better convergence rates. All of it peaked in the dead hours when the office was dark and the parking lot was empty and the only sound was the HVAC system cycling on and off.
"It's API latency," Jay said. He was standing at the whiteboard—the team had eventually gotten one, despite Justin's original resistance—drawing curves with a dry-erase marker. "The LLM providers have lower load between midnight and five AM Pacific. Response times drop. The agents get faster round-trips. Faster round-trips mean more iterations. More iterations mean better convergence."
Navan shook his head. He was sitting cross-legged on the office couch, notebook balanced on his knee. "Latency doesn't explain the quality improvement. Faster round-trips would mean more output, sure. But the satisfaction scores are higher too. The code isn't just more of it—it's better."
"What's your theory?" Jay asked.
"Stochastic resonance."
Jay put the marker down. "You're going to have to explain that."
"In signal processing, stochastic resonance is when adding noise to a system actually improves signal detection. Below a threshold, the signal is too weak to detect. Add noise, and the noise occasionally pushes the signal above the threshold. The system performs better with noise than without it."
"And you think the agents at 3 AM are benefiting from... noise?"
"I think the lower load on the LLM providers changes the distribution of the model's outputs. Slightly different temperature behavior. Slightly different token sampling. It's not better in the obvious sense—it's differently random. And maybe differently random, at 3 AM, is the right kind of noise for the scenarios we're running."
Jay stared at the whiteboard. Navan stared at his notebook. Neither theory was entirely satisfying, which was itself a kind of answer.
They brought the finding to Justin. Justin looked at the charts. He looked at the satisfaction curves. He looked at the hour-by-hour breakdown.
"Don't publish this," he said.
"What?" Jay looked genuinely startled. "This is a real finding. Reproducible. Six weeks of data."
"If we publish it," Justin said, "two things happen. First, every team with an agent-driven workflow shifts their workloads to 3 AM Pacific, which eliminates the advantage. Second, the LLM providers notice the load shift and rebalance their infrastructure, which eliminates the advantage from the other direction."
"So we keep it quiet," Navan said.
"We keep it quiet and we take advantage of it. Schedule the hardest scenarios for the dark hours. Let the agents work when they work best."
Jay looked at the charts one more time. Forty percent. Between midnight and dawn. The agents didn't sleep, but they had a circadian rhythm anyway. It made no sense. It was right there in the data.
"The factory keeps its secrets," Navan said, and wrote something in his notebook, and didn't show anyone what it was.
The stochastic resonance theory is genuinely fascinating. The idea that slightly noisier outputs could improve convergence is counterintuitive but it tracks with simulated annealing and other optimization techniques. Navan reads the right papers.