A few months ago I exported my gradebook as a CSV — three classes, a term of quiz scores, attendance flags, homework submission rates — opened a terminal window, and typed: "Look at these results and tell me which students have declined since March, and whether attendance correlates with it." Thirty seconds later I had an answer, a chart, and a shortlist of six students to speak to on Monday. The tool doing this was not an education product. It was a coding agent, of the kind software engineers have been using daily for a couple of years, and it had quietly become the most useful piece of technology in my teaching life.
Teachers have not been slow to adopt AI, exactly — they have adopted a particular kind of it. Packaged tools like Magic School AI offer friendly web interfaces with buttons labelled "Lesson Plan" and "Rubric Generator", and they do a decent job at the tasks they anticipate. They also make a great deal of money from the gap between what teachers need and what teachers currently know how to ask for. The economics are understandable: if you do not know how to use the raw tools, a packaged one is worth paying for. But the packaged model has real limits. You work through a web form. You upload files one at a time. You hit rate limits just as you get going. And the tool knows nothing about you except what you paste into the box.
Terminal coding agents — Claude Code, Kimi, OpenAI's Codex, and their cousins — work differently, and the differences matter more for teachers than the industry has noticed. They were built for programmers, but a teacher's administrative life is, at bottom, a programming-shaped problem: files, data, text, repetition.
What the terminal actually gives you
The biggest difference is context. A coding agent sits in a folder on your own computer and can read every file in it. There is no uploading to a web interface. My working folder contains the curriculum PDFs, my scheme of work, anonymised assessment data, and notes from the term. When I ask the agent to draft next week's lessons, it is drawing on all of that at once — the actual syllabus wording, the actual spread of results, the actual reading list — rather than on a generic idea of what a Year 11 English lesson looks like. Focused context produces focused output. The same model that writes mediocre lessons through a blank chat box writes genuinely usable ones when it can see the specification and last week's exit tickets.
The second difference is data work. Teachers sit on piles of CSVs — results, attendance, behaviour logs — and mostly eyeball them, because pivot tables are fiddly and nobody taught us statistics. A coding agent treats natural language as the query. "Which classes have the widest spread in the last assessment?" "Flag students whose homework rate dropped below half." "Show me whether my top set is coasting." These are ten-second questions now, and they change what you know about your class before parents' evening.
The third is drafting at scale. Report-card season used to cost me a weekend. Now I point the agent at the anonymised grade data and my comment bank, and it drafts detailed, evidence-referencing comments per student, which I then edit heavily — the agent gets the structure and the evidence right, and I supply the knowledge of the child. It is faster, but more importantly it is better: every comment cites actual data rather than my fading memory of October.
Finally, there is no rate limit in the packaged sense, because these tools are built for people running long, heavy tasks all day. Building a syllabus from a stack of curriculum PDFs — extracting objectives, sequencing them, flagging assessment points — is a half-hour job, not an August.
A necessary caution
Two caveats, stated plainly. First, safeguarding and data protection come before convenience: student data must be anonymised before it goes anywhere near an external model, and your school's policies apply. I work with initials and ID numbers, never names. Second, these tools do not remove the teacher's judgement — they remove the drudgery that crowds judgement out. A report comment the agent drafts is a draft. You remain the editor, and you should.
Software engineers and lawyers rebuilt their workflows around these tools a year or two ago. Teaching will get there, because the underlying shape of the work is the same: too much information, too little time, and a human being at the end who needs your attention rather than your typing. The terminal looks unfriendly for about an hour. After that, it is simply where the leverage lives.