Software Factory Archive

The spinner was the first thing Navan noticed. A small rotating character in the terminal—the classic sequence, | / - \—spinning beside a label that described the current operation. Generating interview questions... Spin. Writing design overview... Spin. Implementing task 3 of 7... Spin.

It was a small thing, but it meant the terminal wasn't silent. When agents were working, you could see them working. Not the actual work—you couldn't watch the tokens stream by or the model think—but the evidence of work, the visible proof that something was happening behind the screen.

The spinner lived in progress.go, alongside the progress bar and the status formatting logic. Navan read the file on a quiet morning, curious about the implementation. It was elegant in the way that good terminal UI code always was—careful about cursor positioning, aware of terminal width, graceful about overwriting lines without leaving visual artifacts.

The progress bar appeared during sprints. A horizontal bar that filled as tasks completed. [=====> ] 3/7 tasks. Each completed task pushed the bar forward. Each push was accompanied by a brief summary: what the task was, how long it took, whether it succeeded or needed a retry.

Below the progress bar, a list of completed tasks with green checkmarks and pending tasks with empty circles. A real-time manifest of work done and work remaining. The kind of display that made you feel like the system was trustworthy, not because it told you it was working, but because it showed you.

Jay appreciated the progress tracking for practical reasons. He could glance at the terminal and know, in one second, where the sprint stood. No need to run agate status in a separate terminal. No need to open the .ai/ directory. The information was there, live, updated with every agent action.

But Navan appreciated it for a different reason. He appreciated the satisfaction of watching agents converge.

There was a specific moment, about two-thirds through a sprint, when the progress bar crossed the halfway point and the momentum became visible. Tasks were completing faster than they had started—partly because the parallel tasks ran simultaneously, partly because later tasks were often simpler than earlier ones. The bar would accelerate, the checkmarks would accumulate, and then the assessment would begin, and you could feel the system approaching the attractor state.

"It's like watching a download bar," Navan told Jay. "Except the thing being downloaded is a piece of software that didn't exist twenty minutes ago."

"That's exactly what it is," Jay said. "progress.go is measuring the distance between where you are and where you want to be. And the gap is closing."

Navan wrote in his notebook: progress.go. Spinner + bar + task list. Convergence you can watch. The gap closes.

The spinner kept spinning. The bar kept filling. The agents kept working. And the humans kept watching, because watching convergence never got old.