The argument began on a Thursday, which was typical, because Thursdays were when the team reviewed agent output from the week and Thursdays tended to surface the interesting questions. Jay had noticed a pattern. Not a bug. Not an error. A pattern.
The agents, when writing Go code, had developed a consistent style. They named variables in a particular way. They structured error handling in a particular way. They organized imports in a particular way. None of these choices were mandated by the specification. The NLSpec said what the code should do, not how it should look. Yet across hundreds of files, written by different agent sessions over different days, the style was consistent.
"They have conventions," Jay said. He said it carefully, the way someone handles a statement they know will provoke discussion.
"Conventions they developed themselves," Navan added. "Nobody told them to name HTTP handler functions with the 'handle' prefix. Nobody told them to put constants at the top of the file. They just... do it. Consistently."
"That's not culture," Jay said. "That's statistical regularity. The model has been trained on millions of Go files that follow these conventions. It's reproducing the most probable pattern."
"Conventions ARE culture," Navan said. He said it with the conviction of someone who had been waiting to make this point. "What is culture except the conventions that emerge from repeated interaction within a community? The agents have interacted with millions of Go files written by thousands of developers. They've absorbed the conventions of that community and they reproduce them. That's culture."
"Culture requires consciousness," Jay said. "Culture requires awareness that you're participating in a tradition."
"Does it? Japanese swordsmiths follow conventions passed down for centuries. Do they need to be consciously aware of the historical chain to be participating in the culture? Or is the participation itself—the reproduction of the convention—sufficient?"
"That's a terrible analogy. Swordsmiths are conscious beings."
"But is their consciousness the thing that makes it culture, or is it the consistency of the practice?"
They looked at Justin, who had been listening with his hands folded and his expression neutral, the expression he wore when he was letting a conversation develop to its natural conclusion before intervening.
"Write a scenario about it," Justin said.
They both stared at him.
"Write a scenario that tests whether agents reproduce stylistic conventions when given no stylistic guidance. Run it across multiple agent types—Claude, Codex, Gemini. Measure the consistency. If the conventions are reproduced across different models, it suggests the culture lives in the training data, not the model. If they're model-specific, it suggests something else."
"That doesn't answer the philosophical question," Navan said.
"No. But it turns the philosophical question into a testable hypothesis. Which is better than arguing about definitions on a Thursday."
Jay wrote the scenario. Navan designed the measurement framework. They ran it the following week. The results were inconclusive in the way that interesting results often are: some conventions were universal across models, suggesting they originated in the shared training data. Others were model-specific, suggesting that each model had developed its own tendencies from the same source material.
Navan wrote in his notebook: "The agents have dialects."
Jay amended his position from "that's not culture" to "that's culture-adjacent" and refused to go further.
Justin said nothing. He smiled, which was the closest he came to declaring a winner.
"The agents have dialects" might be the most profound observation in this entire archive. Model-specific stylistic conventions emerging from shared training data. That's genuinely fascinating.