← Terminal to Agents
Terminal to Agents

An Agent in Your Terminal — and How Much to Let It Touch

18 Sept 2026·7 min read

Everything in the last nine parts is what a coding agent uses to do its job. Not a metaphor — the same commands, in the same shell, with the same permissions your account has. Nothing new gets installed under the hood.

The one difference that matters

Until now you told the machine how. You picked the command, the flags, the file. With an agent you say what you want, and it works out the how — including the parts you did not know were needed.

Compare the two shapes. In Part 6 you wrote: search these files for the word error, and then count the results. To an agent you say: find out why the nightly backup stopped running last Tuesday. The second one is not a longer command. It is a different kind of instruction — it has no fixed steps, and the steps depend on what it finds.

Getting one

npm install -g @anthropic-ai/claude-code
The npm route, which is how the copy on this machine got here. Needs Node installed first.
claude --version
Confirms it is there. Mine printed 2.1.251 while this post was being written.

Then you cd into a project folder and start it there. That is not a detail — Part 3 said the folder you are standing in is where commands happen, and this is where that lands. The folder you start it in is the folder it works in.

What it actually does, in order

  • It reads before it writes. Given a task, it looks at the relevant files first, so it is working with your actual project instead of a guess about projects in general.
  • It runs real commands. The same grep, ls and git from this course — and you can read every one before it goes.
  • It reads the error output and tries again. This is the part a script can never do, and it is only possible because of Part 7: the terminal reports failure in a form a program can read.
  • It stops and asks before doing something it cannot take back. That prompt is not it being unsure — it is the design.

Now the part people skip

An agent in your terminal has exactly your powers. Not more, and that is worth being clear about — but not less either. Everything your account can reach, it can reach: your files, your keys, your network, your ability to delete without a recycle bin.

Three habits, and they are cheapCommit before a big task, so undo is one command — that is why Part 8 taught git. Keep approval switched on for anything that writes, sends or deletes; convenience flags that skip every check exist, and a folder you can rewind is not a reason to use them. And read the command it proposes, not just the summary of it. You spent nine parts learning to read commands. This is the day that pays.

That last one is the honest reason this course started at pwd instead of starting here. Someone who cannot read a command can only approve or refuse on faith. You can look at a proposed line and tell whether it does what was described — which is a completely different relationship with the tool.

Why the terminal and not an app

Because of the point at the end of Part 9. The terminal was already the largest toolbox ever built, and everything in it takes text and gives text back. So an agent that can type inherits all of it immediately — every tool you have installed, and every one you install next week, with nobody writing an integration for either.

Do this in 60 secondsIn a folder you do not mind touching, run git init and git commit -m "before" first. Then start the agent there and give it something small and readable — describe what these files are for. Read what it does. You are looking for one thing only: do you recognise the commands.
Today’s AI angleThe interesting limit is not intelligence, it is reach. An agent in your terminal can only touch what has a command on this machine. Your company ticket tracker, a hosted database, a payments dashboard — none of those has a command sitting on your laptop. That gap is the last two parts.

So what happens when the thing you need is not on your machine at all? That is not a command problem, and the answer has a name.

An Agent in Your Terminal — and How Much to Let It Touch · Ali Huzaifa