Field notes
Learn Claude Code or Codex, not the tool of the week
There is a new AI tool to learn every week, and chasing all of them is a waste of time and money. The better bet is to go deep on one CLI agent, and the two worth your time are Claude Code and Codex. Here is why they are so powerful, and how they work.
Claude Code and Codex are AI coding agents that run inside your terminal. You open a terminal in your project, start the agent, and tell it what you want in plain language. From there it works directly on your machine: it opens your files, changes them, runs commands, and reads what comes back. There is no copy and paste step. The model is driving the computer the same way you would.
That single design choice is the whole story, so it is worth slowing down on. Why does running in the terminal make these tools so much more capable than the same model in a chat window?
Why the terminal is so powerful
Look at what the terminal actually is. It is not a niche tool for hackers. It is the control panel for the entire machine. Almost everything your computer can do has a command behind it: create and edit files, install software, run a program, manage git, query a database, call an API, start a server. The apps you click through are friendly covers sitting on top of commands that already exist underneath.
So when you hand an agent the terminal, you are not handing it one feature. You are handing it the same reach you have. Anything you could type at a command line, it can do too, and on a developer's machine that is close to everything. The model stops being something you talk to and becomes something that operates your computer.
What that lets the agent do
In practice the terminal turns into a concrete set of abilities. The agent reads your whole codebase to learn how the pieces fit. It edits files in place, across as many of them as the change touches. It runs your build and your tests and sees the actual output. It installs the package it needs. It commits the work with git. When a step needs an outside service, it calls the API. None of this is pretended. The commands run for real, and the agent reads the real results.
Chat explains. The terminal does.
This is the line I keep coming back to. A chat app is brilliant at telling you what to do. It hands you the steps and you go run them yourself. A CLI agent skips that handoff. It is standing inside the machine with its hands on the keys, so it just does the steps. Same model underneath, completely different job.
- Tells you the steps
- You copy, paste, and run them
- Lives in a browser tab
- Runs the steps itself
- Edits the files, runs the build
- Lives in your terminal
How they actually work: the loop
The terminal is the capability. The loop is how the agent uses it well. It does not write the code once and hope. It works in a cycle: read the project and your request, make a plan, edit the files, then run a command to check itself, usually the build or the tests. If that command fails, it reads the error output, works out what broke, fixes the file, and runs it again. It keeps going around until the command passes. Watch one full lap below.
Claude Code
Codexfixes the error
That retry is the part you cannot get from a chat window. When a command fails in chat, you are the one reading the stack trace and pasting it back. In the terminal, the agent already has the logs and just keeps going. You are not in the loop. You are reviewing the result of it.
Where to spend your time
This is the real takeaway. There is a new AI tool every week, and trying to keep up with all of them costs real time and money for skills that go stale in a month. A CLI agent is the opposite kind of bet. It sits on top of the terminal, so what you learn keeps paying off no matter which model is ahead this quarter. Put your hours into one of these, not into the tool of the week.
Claude Code and Codex run on the same idea, so you do not need both to start. Pick one, point it at a small real task, and watch it go a few laps around the loop. That is usually the moment it clicks. One agent you know well will take you further than ten you have only skimmed.
In the next post I will get into the layers you need to go from using an agent to building your own.