The question came fifteen minutes into the Q&A, from a woman in the third row wearing a Cognition Labs t-shirt. She didn't look hostile. She looked genuinely curious, which was almost worse.
"How does the StrongDM Software Factory compare to Devin?"
The conference hall was maybe two hundred people, most of them engineers or engineering managers who had come to hear Justin talk about non-interactive development. The talk had gone well. He'd shown the architecture slides, explained the Digital Twin Universe, walked through a live scenario. The audience had been engaged, skeptical in the productive way. Now this.
Justin adjusted the microphone. "So Devin is an AI software engineer," he said. "An autonomous agent that can plan, write code, debug, deploy. It operates in a sandboxed environment, it uses real tools, it works through problems end to end. It's a genuine achievement."
A small murmur. Nobody had expected him to lead with a compliment.
"The Cognition team solved a hard problem," Justin continued. "They proved that an agent can hold context across a complex multi-step engineering task and produce something that works. That's not trivial. Most agents fall apart after three or four steps. Devin doesn't. That matters."
He paused. The audience waited.
"But Devin is an answer to a different question than the one we're asking. Devin asks: can an AI agent do the work of a software engineer? That's a question about capability. Our question is: what happens when you design an entire development methodology around the assumption that AI agents are the only ones writing and reviewing code? That's a question about process."
He took a sip of water. "Devin is a very smart contractor. You give it a task, it does the task. The factory isn't a contractor. It's an ecosystem. The agents don't just write code—they validate it against behavioral clones of real-world services, they measure probabilistic satisfaction metrics, they generate and expand test scenarios. The humans don't review the code because the system is designed so that code review is the wrong bottleneck to optimize for. Instead, we optimize for fidelity of the simulated world the code runs in."
"So you don't think they're competitors?" the woman asked.
"I think they're convergent evolution," Justin said. "Two species arriving at similar body plans from different starting points. Devin started from the question: what if an AI could code like a human? We started from the question: what if humans stopped coding entirely and focused on describing the world the code has to survive in? Both approaches work. Both are real. The interesting thing is what happens when they meet in the middle."
After the talk, the woman from Cognition found Justin at the reception. They talked for forty minutes about agent evaluation frameworks. She showed him a benchmark result he hadn't seen. He showed her a satisfaction curve from a scenario that had taken three weeks to converge.
"You were generous up there," she said.
"There aren't enough of us doing this work to waste time being competitive about it," Justin said. "We're all trying to find the same cliff and jump off it. We just picked different run-up angles."
She laughed. "That's either very wise or very reckless."
"In my experience," Justin said, "those are the same thing."
"Convergent evolution" is such a perfect framing. Not competing, not collaborating, just two different lineages arriving at the same future from different starting points. This is the most quotable fic in the archive.