AI Coding Agents: Terminal vs Editor
There's a quiet shift happening: a lot of AI-assisted development is moving out of the editor and into the terminal. Here's why, and when each setup makes sense.
The editor approach
Inline assistants like GitHub Copilot shine at autocomplete and small, context-aware edits. They see your open files and diff, which makes them excellent for incremental work. The downside is they operate within the editor's view of the world — their context is bounded by what's on screen and in the index.
The terminal agent approach
Terminal agents run as first-class tools. They can execute commands, run tests, edit files, and explore an entire repository on their own. Instead of suggesting a fix, they make the change and verify it. This "agentic" loop is closer to pairing with a careful junior developer than to autocomplete.
- Runs anywhere the terminal runs — servers, containers, remote machines.
- Plays nicely with toolchains that are already CLI-first (git, test runners, formatters).
- Model choice is usually flexible — bring your own provider or run local models.
Which should you use?
They complement each other. Widely used pattern: let the editor handle fast inline help, and hand larger "make this work and prove it" tasks to an agent. Many people keep both and reach for the right one per task.
My current setup
I keep an inline assistant for autocomplete and a terminal agent for anything involving multiple files, tests, or unfamiliar code. The terminal agent's ability to actually execute and verify changes is what won me over.