The Harness vs. the Interface

Wed, 14 Oct 2026 09:00:00 -0600

There’s a debate around AI coding tools that revolves around screen real estate. Be it on Reddit, HackerNews, LinkedIn, or your own development team, we’ve argued over whether a developer belonged in a chat sidebar in VS Code, in an interactive terminal session using a CLI/TUI tool, or in some specialized graphical desktop tool.

Each setup shares the same fundamental premise. They treat software development as an ongoing, turn-by-turn chat between a human and a model. You write a prompt, the model streams tokens back, you inspect the diff, and you approve the file write.

That back-and-forth makes sense for small utility functions (or when you genuinely want to enjoy the love that is regex), but when you ask an agent to tackle substantial work, like modernizing an API client across forty files or running a multi-step test suite, the conversation model collapses into window juggling. You end up managing half a dozen terminal windows, switching contexts constantly, and babysitting intermediate tool outputs.

The friction isn’t about choosing between a terminal or an editor window. It comes from coupling the execution engine directly to an interactive chat prompt.

Git-Worktrees on David R. Longnecker - Converting Coffee to Code

The Harness vs. the Interface