The first inflection was October 2024. Claude 3.5, second revision. The moment the agents crossed from assistants to workers. Justin had named it, documented it, built a factory on top of it.
The second inflection was November 2025. And this time, nobody needed to name it. Everyone felt it.
Claude Opus 4.5 and GPT 5.2 landed within ten days of each other. Jay noticed first because Jay was the one who watched the metrics most obsessively, the way a sailor watches the horizon. He came into the office on a Wednesday morning, opened the satisfaction dashboard, and stopped moving.
"Justin," he said. His voice was even, the way voices get when someone is working very hard to stay calm. "Come look at this."
The satisfaction curve, which had been holding steady in the high 97s for months—good, excellent even, a number the team was proud of—had jumped to 99.2 overnight. Not because anyone changed the scenarios. Not because anyone refined the specs. The agents had simply gotten better.
Justin stood behind Jay's monitor and looked at the curve. Then he looked at the date. Then he looked at the model versions in the agent configuration.
"They upgraded the underlying model in the Anthropic API," Justin said. "Opus 4.5 went live yesterday."
"And OpenAI pushed 5.2 last week," Jay added. "The Codex agents are using it already."
Navan was pulling up the detailed metrics. He had the breakdown by scenario category, by twin interaction, by complexity tier. Everything was up. Not uniformly—the simple scenarios showed modest improvement, which made sense because they were already near the ceiling. But the complex scenarios, the ones involving multi-service orchestration and edge-case handling, had jumped dramatically.
"The hard problems got easier," Navan said. He was writing numbers in his notebook, fast. "The Jira-to-Slack-to-Drive provisioning chain that used to converge at 94 percent? It's at 98.7 now. Same scenarios. Same twins. Different agents."
Justin was quiet for a long time. When Justin was quiet, the team had learned to wait.
"We need to recalibrate everything," he finally said. "The scenarios that were testing edge cases are now too easy. The satisfaction thresholds we set in August are too low. The curve bent, and everything downstream of the curve needs to adjust."
"That's a lot of work," Jay said.
"That's the best kind of work. Our tools got better and now we get to raise the bar."
They spent the next two weeks rewriting scenarios. Making them harder. More edge cases. More complex orchestrations. More adversarial conditions. The Digital Twins got meaner—more realistic failure modes, more latency variation, more of the chaos that real production environments inflicted on software.
By mid-December, the satisfaction metric had settled at 98.1 percent against the new, harder scenarios. Lower than the 99.2 they'd briefly celebrated. Higher than the 97.6 they'd been proud of before. The number told a story: the agents were dramatically better, and the team had responded by demanding dramatically more.
Justin wrote in his internal document: The second inflection. The curve bends again. We recalibrate. This is the rhythm now.
He didn't say how many more inflections there would be. Nobody knew. The curve was still bending.
The detail about them responding to better agents by writing HARDER scenarios is so perfectly in character for this team. They don't celebrate. They recalibrate. It's both admirable and slightly unhinged.