CLI AI tools like Claude Code, Cursor CLI, Copilot CLI, OpenAI Codex CLI, Google Gemini CLI, OpenCode, and so on, are now my primary means of writing code. Instead of having an editor at the center of my screen, and other tools such as terminals and AI agents that augment this setup, now, the AI agent is at the center of the screen, and my editor and terminal augment my flow with the AI agent.
The basic idea of this is that because the AI agent benefits from accurate context, it now makes sense to perform all actions through the AI agent, so that it maintains accurate context as we go. Whereas if I use my editor or terminal for a whole bunch of tasks, those tasks are not reflected in the context window of the AI agent, so that it has less knowledge about what’s going on. However, if I do all these tasks through the AI agent, such as moving files, renaming files, renaming variables, adding and subtracting code, running tests, running the type checker and the linter, etc., through the AI agent, then all of that context is in the AI agent so that it can help perform whatever is the next step more accurately.
It’s worth noting that my flow for writing software is still fundamentally the
same. Although the AI is my IDE now, I still have to have clear product
requirements and clear technical requirements. It is my responsibility to
communicate, primarily in English, but also in code samples, what these
requirements are to the AI agent, and to constantly look over the AI agent to
make sure that it is actually following the rules. If it’s not, this may be a
sign that documentation such as in the CLAUDE.md file or AGENTS.md file may
need to be updated.
This new paradigm makes it possible to rely on AI agents to a very large extent when creating or updating an application. The latest AI coding agents are able to perform nearly all tasks that a human programmer would do, including adding new code and tests, and updating existing code, fixing bugs, refactoring, etc. It costs money to use these tools, as much as $100 per day or so for one AI agent, but this cost makes a lot of sense for any commercial code. If you are being paid more than it costs to use the AI agent for these tasks then it makes sense.
AI agents can be seen as a very powerful IDE that has a greater understanding of the code and a greater ability to automate updating the code than prior generation non-AI IDEs. For instance, my normal IDE (Neovim with a bunch of plugins) can do things like rename a variable using the TypeScript language server. This is a great feature. I also have tons of other features such as seeing the documentation on “hover”, and “completions” for things like possible functions or variable names. These are great features of modern IDEs. However, AI provides a different set of features that are far more powerful, including the ability to add new blocks of code that conform to style and architecture guidelines, with tests. This amplifies the sheer volume of code that a human can output.
We are in a disruptive period for software engineering. The new AI agent tools are so powerful, they can become your primary IDE. Instead of going to your IDE/editor by default, you can now go to the AI agent by default. You can still use your editor to supplement the AI, such as reviewing all the files it creates or updates as it goes, or to make the occasional manual adjustments to some of its code. However, by and large, it now makes sense to go through the AI agent for all coding tasks so that the context window stays accurate, so that it knows what’s going on and can use that context for future code changes.