The CLI/TUI split in the article is right, but to me the more interesting boundary now is human-facing vs agent-facing interfaces. For most things TUIs lose on both sides.

We run an internal agentic development harness we experimentally built for working on large-ish projects. first version started as a TUI, but it quickly became unusable for the large scale projects it was meant to work on (essentially PRD to PR, with planning loops, task decomposition loops and dependency aware execution / review loops). The TUI was too state-bound for another agent to drive reliably, and too cramped for a human to review at scale. We ended up replacing the TUI with a headless cli, and every interface we've built for the harness has become either a text artifact (mostly for agents to work) or a web view (for humans to drive and review).