It started as a monitoring tool. Jay needed to track token spend—they were burning over a thousand dollars a day per engineer on API calls to Claude, and the invoices were getting complex enough to require their own infrastructure. So he described what he wanted to an agent: a real-time dashboard. Token consumption by model, by task type, by time window. Running totals. Burn rate projections.
The agent built it in an afternoon. A clean web interface with dark backgrounds and bright data lines, the kind of thing that looked like mission control if you squinted. Jay put it on the wall monitor in the main room, the one they usually used for terminal output.
That was the first day. By the end of the week, it had grown.
Jay added satisfaction curves—the probabilistic validation metrics for every active scenario, plotted over time. You could watch them converge. Some shot up quickly, hitting 0.95 within a few agent iterations. Others climbed in slow, grudging increments, 0.71 to 0.73 to 0.74, like a patient learning to walk again. A few plateaued stubbornly around 0.6 and sat there, mocking everyone.
Then he added convergence rates: how quickly each agent run was reaching stable outputs. He color-coded them. Green for fast convergence. Yellow for moderate. Red for runs that were still oscillating after fifty iterations. The reds were the interesting ones. Something about those scenarios was making the agents uncertain, and the dashboard made that uncertainty visible.
Navan was the first to start watching it the way you watch a fire. He'd pull his chair up to the monitor during lunch and just observe the curves, eating a sandwich one-handed while the satisfaction metrics ticked upward. "The Jira provisioning scenario just crossed 0.9," he'd say to nobody in particular. Or: "Look at that red one. It's been stuck at 0.62 for three days."
Justin noticed it next. He started coming into the room and glancing at the dashboard before he looked at anything else, the way you check the weather before you step outside. The dashboard had become the team's ambient awareness of the factory's health. Its heartbeat.
"You built a campfire," Justin told Jay one evening, when the three of them were sitting around the monitor watching a particularly stubborn scenario finally break through 0.85.
"I built a monitoring tool," Jay corrected.
"You built a thing the team gathers around. A thing that tells us stories. Look—" He pointed at a satisfaction curve that had dipped suddenly and then recovered. "That's a narrative. Something went wrong and then something went right. We don't even need to look at the logs to know that. The dashboard tells us."
Jay considered this. On the screen, three red scenarios shifted to yellow simultaneously. Navan made a small, satisfied sound.
"Campfire," Jay conceded.
He added one more feature that night: a subtle animation that played whenever a scenario's satisfaction metric crossed 0.95 for the first time. Nothing flashy. Just a brief golden pulse around the data point, like an ember catching.
Nobody asked him to add it. He never described it to an agent. He wrote those twelve lines of JavaScript himself, by hand, at 11 PM, and told no one.
Some things the humans have to do.
The last line DESTROYED me. In a factory where the core rule is "code must not be written by humans," Jay writes twelve lines of JavaScript by hand because the campfire needed one more ember. That's not a violation of the rule. That's the exception that proves it.