CLI and Programming Study in the Age of AI Coding Agents [August 2026]: 11 Services, Starting from the Terminal and Git

Claude Code, Codex, Cursor, GitHub Copilot — by 2026, AI agents handle a substantial share of the work of writing code. What has happened is not that "learning to program became unnecessary," but that the contents of the necessary foundation were swapped out.

Agents write code. But launching the agent, reading the diff it proposes, interpreting a failed command, and restoring a broken state are all things a human does in the terminal. Every major coding agent takes the CLI as its main ground.

This article compares 11 learning services, Japanese and international, from one angle: programming and CLI study as the prerequisite for using AI agents well. Pricing, format, language support and free tiers all reflect official figures as of August 2026.

The short version

Why the terminal matters more, not less, as you delegate

Every AI coding agent follows a similar execution model. A human states the goal, the agent reads and writes files and runs commands, and the human approves the result — and the place where that approval happens is the terminal.

Three points matter most.

The 11 services at a glance

Prices include tax and reflect August 2026. Overseas services show official figures (US dollars, or the yen price displayed from Japan).

How to choose: four axes

1. Enter through the language, or through the environment

Most services start from syntax. But when you plan to work with AI agents, running the environment (terminal, Git) first often gets you moving faster: the agent fills in syntax, while operating the environment is not delegated.

For the environment, WebTerm and WebTerm Learn; for the language, Progate, dotinstall or paiza Learning. ZEN Study sits between the two, since Linux, Git, GitHub and Docker are built into its introductory course.

2. Overview by video, or work it out by hand

3. Does it work in your language

Boot.dev, KodeKloud and DeepLearning.AI have no Japanese interface. Reading technical English is itself a professional skill, but starting your very first course in a second language raises the odds of quitting. Build the foundation in your own language, then move across.

4. How far the free tier goes

Free tiers vary widely. WebTerm and WebTerm Learn are entirely free, and DeepLearning.AI's short courses enroll for free. paiza Learning's free tier includes 928 courses and 1,209 exercises, and Recursion offers about 30 hours. ZEN Study, by contrast, is subscription-only with no free study tier.

Recommendations by situation

The order to learn in

Skip step 1 and step 3 tends to end in "it worked, but I have no idea what happened."

The services in detail

Here is each service through one lens: what part of the work it covers when you are using AI agents. The order follows the learning sequence — environment (terminal, Git), then language basics, then implementation and evaluation.

WebTerm | A free learning terminal that runs in the browser

An in-browser Linux terminal implementing over 100 commands: pwd, ls, cd through grep, find, chmod, pipes and redirection, Vim and Git, on an environment with real filesystem state. Nothing to install, no account to create — you can type your first command the moment the page loads.

The core is 15 tutorials (4 on the terminal, 2 on Vim, 3 on Git, 6 on Git troubleshooting), with mode 16 as a free-play sandbox. Because grading looks at the state of the filesystem after your command runs, copying an answer will not move you forward.

What matters most for agent work is the six Git troubleshooting tutorials: undoing commits, resolving merge conflicts, recovering lost commits, switching branches under pressure. These are exactly the operations you need after handing an agent a long stretch of work.

WebTerm Learn | The terminal and Git across 12 courses and 129 lessons

Illustrated slides alternate with exercises across the terminal (introduction 20 / fundamentals 13 / advanced 25) and Git (introduction 23 / fundamentals 18), each stage followed by training courses (3 lessons × 5). The advanced terminal course reaches shell customization, process management and SSH, so you can carry on to the point of working on a server.

Structurally, the notable choice is 41 lessons on Git. As agents take over implementation, what remains on the human side is reading diffs, isolating branches and undoing the wrong commit — all Git operations.

There are also 9 lessons on Vim and, unusually, 6 on the terminal multiplexer Herdr, a multiplexer built for coding agents. Understanding panes, tabs and workspaces first means you will not trip when you start running several agents in parallel.

Progate | Claude Code Introduction (beta), in Japanese

Illustration-led slides paired with in-browser exercises. 3.7 million users across more than 100 countries. Alongside 15+ languages, it carries environment lessons: Command Line (1 lesson + 9 articles) and Git (1 lesson + 9 articles).

What is close to unique in this category is that AI agent courses sit on the same shelf as the language primers. Generative AI Basics has you converse with an AI in the browser to internalize prompt patterns; Claude Code Introduction (beta) covers agent basics, project configuration via CLAUDE.md, automation through skills and safety management via permissions.

Covering permissions in an introductory course is a practical choice. What an agent may and may not do is written in terms of commands, so you cannot scope it without knowing what rm or git push does. Environment knowledge and agent configuration are continuous — the argument of this article, turned into a curriculum.

ドットインストール | dotinstall: 576 lessons in 8,618 three-minute videos

A video-first service running since 2011. Three-minute units keep the cost of starting low — one video per commute — and the lessons play on a phone.

For agent work, the value is that the environment side — local setup, Linux commands, Git — is covered in Japanese. Agents run in your own environment, so getting stuck on Node.js versions or PATH means never reaching the point of launching one. As of August 2026, the site announced that AI-related lessons were being added.

paizaラーニング | paiza Learning: where the exit leads to job listings

Code runs in the browser, so no setup is required, and the free tier alone includes 928 courses and 1,209 exercises — enough to reach working knowledge of a language at zero cost.

What separates it: skill-check ranks (S through E) function directly as job application credentials. The curriculum itself is designed against more than 20 million skill-check attempts and job data from 3,000+ companies, so "learn → measure → apply" completes inside one service.

ZEN Study | Linux, Git and Docker inside the introductory course

Material written from scratch by working Dowango engineers, with unusual breadth for ¥1,100 per month: the same subscription also covers university-entrance preparation, web design and video creation.

The detail worth noting is the learning elements of "Introduction to Programming: Web Application Development" — Linux, Git, GitHub, Docker, web security and Hono beside JavaScript. Environment topics fall inside a course meant for complete beginners, reflecting a stated policy of teaching from basic PC operation without skipping steps.

Beyond it lie large-scale web development in Scala and computer science courses — data structures and algorithms, automata, networking. Environment through CS theory on one subscription has no equivalent at this price.

Techpit | After syntax, ship one whole service

A marketplace of tutorials by working engineers. From 100+ titles you can select backwards from the finished product — "build a matching app in Rails" — which fills the gap after syntax, when the question becomes what to build.

There is no in-browser runtime; you write the code in your own environment, so terminal skills and environment setup are prerequisites. If that is where you stall, running WebTerm Learn or Progate's Command Line lessons first changes how it goes. INTLOOP Inc. acquired the business in November 2025 and it now runs at techpit.intloop.com.

Recursion | CS fundamentals in Japanese, designed by ex-Meta engineers

800+ hours selected from US university CS curricula and built at a ratio of 20% input, 80% output. It covers data structures and algorithms, OOP, databases, networking, operating systems and system design in Japanese, with 100+ backend assignments that avoid frameworks.

The more implementation agents take on, the more the human job becomes evaluating what comes out. Estimating complexity, choosing a data structure, spotting a table design that will not hold — these belong to the person reading generated code, and that is exactly the layer Recursion covers. Its stated goal, "escaping copy-paste engineering," means more in an AI-first environment, not less.

Boot.dev | Backend and DevOps in one path, with an AI tutor

A challenge-based service from Ogdolo LLC. Python, Go, TypeScript and SQL sit beside Linux, Git, Docker, Kubernetes, CI/CD and AWS, so the familiar gap between "I can write it" and "I cannot deploy it" rarely opens. More than 1.2 million learners.

Its AI tutor, Boots, answers questions with Socratic questions rather than solutions, steering you to locate the hole in your own understanding. Combined with a format that requires working code, it is designed to produce people who can evaluate AI output rather than depend on it.

KodeKloud | Real labs and an AI path with 8h32m on Claude Code

A hands-on platform for DevOps, cloud and AI from Zaurac Technologies (Singapore). 80+ courses and 880+ labs, with environments for Kubernetes, Docker, Terraform and more that start instantly. The motto — "Build. Break Fix. Learn." — describes the design.

It is the only service here carrying an AI learning path (19 courses, ~106 hours) as a formal curriculum, running from Prompt Engineering 101 (5h) and MCP For Beginners (2h31m) through Claude Code For Beginners (8h32m) and Cursor AI (3h48m) to MLOps and AIOps.

An eight-and-a-half-hour Claude Code course sitting on the same shelf as Linux, Docker and Kubernetes says it plainly: a coding agent is one development tool among others, taught with the environment knowledge it depends on.

DeepLearning.AI | The Anthropic co-produced Claude Code course, free

The AI learning platform founded by Andrew Ng, publishing 104 short courses that finish in one to three hours — most of them co-produced with the companies that build the tools. Courses on features that shipped weeks earlier keep arriving, so the material tracks current behavior.

For coding agents there are Claude Code: A Highly Agentic Coding Assistant (2 hours, intermediate, with Anthropic) and AI Coding Workflows: From Cloud to Local (1h32m, intermediate, with JetBrains), both listed as "Enroll for Free." Forty-three agent-related courses reach as far as designing multi-agent systems with crewAI.

The prerequisite is being able to read code. Put the other way: if you can, the closest thing to official material is available at zero cost.

FAQ

If AI writes the code, why learn to program at all?

The share of work spent writing has certainly fallen. But reading, evaluating and recovering remain human work. A person approves the diff an agent produces, and as long as responsibility comes with that approval, being able to read is a requirement. It is closer to the truth to say the target moved from "writing from scratch" to "reading and judging."

Terminal first, or programming first?

If working with AI agents is the premise, do the terminal and Git first. One or two weeks is enough for the basics, and afterwards no language you pick will strand you at environment setup. WebTerm and WebTerm Learn are both free.

How far can I get for free?

Terminal and Git fundamentals are fully covered for free (WebTerm, WebTerm Learn). Language syntax goes a long way on paiza Learning's free tier (928 courses, 1,209 exercises). Even agent usage is free through DeepLearning.AI's short courses. Paying starts to matter at systematic CS fundamentals (Recursion) and real infrastructure practice (KodeKloud).

Should I avoid English-only services?

Make your first course one in your own language. That said, Boot.dev, KodeKloud and DeepLearning.AI hold ground that Japanese services do not — real-environment DevOps practice and vendor-official agent courses. Build the base first, then cross over when you need to.

Conclusion

AI coding agents did not remove the need to learn; they changed the order. Rather than perfecting syntax before implementing anything, it is now faster to secure the environment (terminal, Git) first, delegate to an agent while reading the diffs yourself, and fill in the knowledge required for evaluation afterwards.

The first step costs nothing. Type commands in WebTerm, structure it with WebTerm Learn, and move on to Progate or Recursion — a detour that turns out to be the shortest route.

For choosing the agent itself, see "Comparing AI coding tools"; for using Claude Code, "The complete Claude Code guide". For learning services compared by purpose, see "11 programming learning services compared".

This article reflects information published on each company's official site as of August 2026. Pricing and course structure change; check the official sites for the latest details.