The cause? iTerm's scrollback buffer was set to "Unlimited." I never thought twice about it. I close terminal windows. I move on.

But AI coding agents don't close windows. They don't move on. They generate thousands of lines of output all night. iTerm stored every byte.

One default setting I'd never questioned crashed my machine overnight.

"Unlimited" was perfectly reasonable. Nobody designed it thinking a tireless agent would be generating output at 3am.

The setting didn't break. The assumption behind it did.

This pattern is everywhere now. AI makes it easy to generate content, code, analyses. But how we review, ship, and decide? Same as before.

MIT's 2025 "GenAI Divide" research: 95% of enterprise GenAI pilots deliver zero P&L impact. The barrier isn't the models.

It's that most orgs haven't changed how they actually work around what AI makes possible.

Teams spin up agents, get dazzled by output, ship it without changing how they review or decide. Budgets evaporate. Nothing ships faster.

It's the scrollback problem at org scale. The defaults were fine before. They're not anymore.

The trough of disillusionment isn't about AI underperforming. It's about everything around AI not keeping up.