The bug report came from an agent. Not from a human filing a ticket, not from a monitoring alert, but from an agent running a scenario against the Slack twin that had encountered a rendering discrepancy in a Block Kit message. The agent described the problem with clinical precision: a section block containing an image accessory and a markdown text element was rendering with incorrect padding when the image was wider than 320 pixels.
Nobody at StrongDM should have been able to diagnose this from the description alone. Block Kit was Slack's layout framework—a JSON-based system for building rich messages with buttons, images, text fields, and interactive elements. It was well-documented but deeply idiosyncratic. The rendering behavior depended on the client (desktop, mobile, web), the block type, the element combination, and sometimes the phase of the moon.
Navan read the bug report and said, "The image accessory in a section block has a maximum display width that varies by client. The twin is using the web client's rendering rules, but the scenario is assuming the desktop client's behavior. The padding calculation differs by four pixels."
Jay stared at him. "You knew that from memory?"
"I built the Slack twin. I had to understand every Block Kit layout combination to make the twin render messages correctly. Some of these behaviors aren't in the documentation. I had to test them empirically—send messages with specific block combinations and measure the rendered output pixel by pixel."
"Pixel by pixel?"
"The twin has to produce responses that are behaviorally identical to real Slack. If a scenario sends a Block Kit message and the rendering is off by four pixels, that's a divergence. It might not matter for most use cases, but we don't get to decide which divergences matter. The scenarios decide."
Justin, who had been listening from the doorway, said, "You probably know Block Kit better than most people at Slack."
Navan looked uncomfortable with the compliment, in the way he always looked uncomfortable with compliments—not falsely modest, just genuinely uncertain about how to receive them. "I know the rendering behavior. The actual Slack engineers know the implementation. Those are different kinds of knowledge."
"But complementary," Justin said. "You know what the system does. They know how it does it. For building twins, your kind of knowledge is more useful."
This was the strange inversion of the factory. Traditional engineering valued implementation knowledge—how the code worked, how the database was structured, how the API handled errors internally. The factory valued behavioral knowledge—what the system did from the outside, what it looked like to a client, what patterns emerged from thousands of interactions.
Navan had accumulated this knowledge through months of painstaking observation. He'd sent thousands of Block Kit messages through real Slack and compared the results against his twin's output. He'd catalogued undocumented behaviors, edge cases, client-specific rendering differences. He'd become an expert in something that had no certification, no textbook, no formal body of knowledge. He'd become an expert in what Slack actually did, as opposed to what it said it did.
The four-pixel padding discrepancy was fixed by the agent in eleven minutes. Navan's knowledge of why it existed had taken three months to acquire. Some expertise can't be hurried.
As someone who has fought with Block Kit rendering inconsistencies across clients, the pixel-by-pixel testing is not exaggeration. It really does vary between desktop and web. The fact that Navan mapped all of this for a twin is extraordinary.