Working with Coding Agents
Ship software with an AI pair that does the typing
Use Claude Code, Codex, Cursor, and similar tools to build real software. Write tasks agents can finish, review their work, and run them safely with git.
- Lessons
- 7
- Reading time
- 68 min
- Level
- Beginner
- Free lessons
- 1
Lessons in this module
- 2.110 minFree
The Coding Agent Landscape
A tour of terminal agents, IDE agents, and cloud agents — Claude Code, Codex, Cursor, and more — and how to choose between them.
- 2.29 min
Writing Tasks an Agent Can Actually Finish
The anatomy of a great agent task: goal, context, constraints, and a definition of done. Includes before/after rewrites of real prompts.
- 2.39 min
Project Instructions: CLAUDE.md and AGENTS.md
Teach an agent your codebase once. What to put in project instruction files, what to leave out, and templates you can copy.
- 2.48 min
Plan First, Then Execute
Why planning mode prevents most agent disasters, how to review a plan, and how to break large features into agent-sized steps.
- 2.58 min
Reviewing Agent-Written Code
A fast, reliable review process for code you didn't type: tests first, diff reading strategies, and red flags to catch.
- 2.68 min
Git, Branches, and Worktrees for Agent Work
Use git as your safety net: small commits, branches per task, worktrees for parallel agents, and easy rollbacks.
- 2.716 min
Building a Full App Agentically, Start to Finish
A complete walkthrough of building and deploying a real web app with a coding agent — from blank folder to live URL.
About this module
A coding agent is an AI agent that works inside a software project. Instead of suggesting a single line of code, it reads files, searches the codebase, edits several files at once, runs tests and other commands, and checks its own results, all from a plain-language request. Coding agents come in three main shapes: terminal agents such as Claude Code, Codex CLI, and Gemini CLI; agents built into editors, such as Cursor and GitHub Copilot; and cloud agents that work on a repository in the background and hand back a pull request.
These tools can now handle substantial work, but results depend heavily on how you use them. Vague requests produce code that misses the point, missing project context produces code that ignores your conventions, and unreviewed changes can hide subtle bugs. The skills that fix this transfer across every tool: writing clear tasks, giving the agent context it can't find on its own, planning before executing, and verifying the result.
This part teaches those skills hands-on. You will survey the coding agent landscape and pick a tool, learn the four parts of a task an agent can finish, and write project instruction files such as CLAUDE.md and AGENTS.md so the agent understands your codebase from the first message. You will use plan mode to catch bad approaches early and follow a fast process for reviewing AI-generated code that starts from tests rather than reading every line.
You will also use git as a safety net, with small commits, a branch per task, a rollback cheat sheet, and git worktrees for running several agents in parallel. The part ends with a complete walkthrough of building and deploying a small web app with a coding agent, from an empty folder and a written spec to a live URL.
Questions about Working with Coding Agents
What is the best AI coding agent?
There is no single best one; the leading tools are close in capability and change quickly. Terminal agents like Claude Code and Codex suit people comfortable on the command line, editor-based agents like Cursor and GitHub Copilot suit people who want to stay in their IDE, and cloud agents suit background work that ends in a pull request. Pick one and learn it well, since the core skills transfer.
What is the difference between Claude Code, Codex, and Cursor?
Claude Code is Anthropic's coding agent, used mainly from the terminal and also available in IDEs and other apps. Codex is OpenAI's coding agent, available as a CLI, an IDE extension, and a cloud agent. Cursor is an AI-focused code editor with a built-in agent that can use models from several providers.
What is a CLAUDE.md file?
CLAUDE.md is a Markdown file in your repository that Claude Code reads automatically at the start of each session. It holds the commands, conventions, gotchas, and definition of done for your project. Many other agents, including Codex and Cursor, read a similar file called AGENTS.md.
Can AI coding agents build a full app?
They can build and deploy a small, well-specified app with you directing the work, as the walkthrough in this part shows. You still decide what to build, break the work into steps, and check that each step is correct. Larger, long-lived codebases need the same discipline applied consistently over time.
Is code written by AI agents safe to ship?
Only after review. Agent-written code can contain subtle bugs, weakened tests, or changes beyond what you asked for. Run the tests and checks yourself, read the diff in a deliberate order, and treat the code as your responsibility, just as if you had typed it.
How do you run multiple coding agents at once?
Give each agent its own isolated copy of the repository, most commonly a git worktree on its own branch, so the agents can't overwrite each other's files. Keep the tasks independent, then review and merge the branches one at a time.