The experiment was Navan's idea. He proposed it on a Monday, and by Wednesday afternoon he was sitting at his desk with a split terminal: the left pane showing his text editor, the right pane showing an agent session running through Leash. The task was to build a new scenario validator for the Slack digital twin, and they were going to build it together. In real time. Human and agent, working on the same problem simultaneously.
The rules were simple. Navan would write scenarios describing the expected behavior. The agent would write code implementing that behavior. They would work in parallel, communicating through the shared CXDB context. Every scenario Navan wrote would be immediately visible to the agent. Every line of code the agent produced would be immediately testable against the scenarios.
Jay pulled up a chair to watch. Justin stood behind them, arms crossed, saying nothing.
Navan started typing. His first scenario described a user sending a message to a Slack channel and the twin correctly echoing back the message with proper timestamps, threading metadata, and user attribution. He saved it. The CXDB context updated. On the right pane, the agent began writing code within seconds.
The agent's first implementation was rough. It handled the basic case but missed edge cases around thread replies. Navan, watching the code appear, immediately wrote a new scenario covering threaded replies. He didn't write it as a correction. He wrote it as a new expectation, an expansion of the behavioral surface. The agent saw the new scenario and adjusted its implementation before Navan had finished typing the third scenario.
"It's reading my scenarios as fast as I write them," Navan said. His voice had an edge of wonder that he was trying to suppress and failing.
"Faster," Jay said. "It started adjusting the implementation before you finished scenario three. It predicted where you were going based on the pattern of scenarios one and two."
The rhythm was like nothing any of them had experienced. It wasn't pair programming in the traditional sense—two humans taking turns at a keyboard. It was something new. Navan described the world. The agent built the world. Navan refined the description. The agent refined the construction. The feedback loop was measured in seconds, not minutes or hours.
By the fortieth minute, the Slack twin's scenario validator was handling message sends, thread replies, emoji reactions, file uploads, channel joins, channel leaves, and the peculiar edge case where a user sends a message to a channel they've been removed from. Navan hadn't planned to test that last case. The agent had added a code path for it after inferring the possibility from the channel-leave scenario.
"It extrapolated," Navan said.
"It extrapolated correctly," Justin amended.
The session lasted ninety minutes. When it was over, Navan's scenario suite had thirty-two entries and the agent's implementation passed all of them. Jay ran the satisfaction metric. It was 0.94. For a component built in ninety minutes, that was extraordinary.
Navan sat back. His hands were trembling slightly, the way they did after he played music for a long stretch. The creative shaking. The kind that meant something had been channeled through him.
"That was electric," he said quietly.
Nobody disagreed. Jay saved the session log. Justin asked them to do it again next week with a harder problem. Navan opened his notebook and wrote a single line: "The agent is not a tool. The agent is a collaborator that doesn't need to rest."
Twenty years of pair programming and I never considered that the other half of the pair could be an agent writing code while I write intent. This reframes everything.