How to start coding with AI agents

There have been two defining moments in AI adoption.

The first was ChatGPT's release in late 2022, when millions discovered they could talk to a machine that actually understood them.

The second is happening right now: autonomous coding agents like Claude Code are showing developers what it means to have AI that doesn't just suggest code. AI writes, runs, and ships it.

What is "agentic coding" anyway?

Agentic coding means you're talking with AI through a chat UI and seeing it implementing and running code for you. Just like if you hired a developer who works for you and livestreams all the results. Agents have access to your system and can run bash commands for you, such as "npm install", "git push", or "rm -rf /".

Previously, AI was mostly used as a tab-completion tool for coding. But in the last months of 2025, LLM models became much better and more powerful, and now can produce great code and are super useful to iterate faster and build quality products.

Now, you can actually build a production-level system if you approach “vibe-coding” as engineering task, but not as some slop beast that is out of control. World-class engineers code with AI: Linus Torvalds (Linux Creator), DHH (Ruby on Rails creator / Basecamp founder), Dan Abramov (core React contributor), Gergely Orosz (running the large engineering newsletter) and so many others.

Quality depends on you 🫵

AI won’t magically write a project for you right away (yet). You need to carefully configure it.

The style and output quality depend on you, as you need to craft your agent’s context to create exactly what you need. This is hard to do if you don’t follow a system in your code, but luckily you can use open-source configurations from others, e.g. React Best Practices from Vercel. (View global directory of skills)

If you don’t have any rules or prompt system, chances are you’ll end up with slop. But now it’s on you, not AI.

Slop — badly produced, generic AI output, that is not refined by a human.

Engineer the Context

Context engineering is the process of loading information into your AI agents. The thing is: LLMs are limited in the chat's context size that they can process, and their output becomes much worse if you use it unwisely. Also, it costs money.

The art of engineering today is the art of creating the best context architectures and writing thoughtful rules. Learn to write less code, but start writing more context-enhancing documents on how to write code.

Use Skills, Rules, and Commands to automate common actions and enforce coding guidelines. Write down rules in Markdown as soon as you repeat something twice to an agent. This is a new DRY.

Subagents let you automate multiple roles, e.g. front-end, back-end or tech writer. You can tag a specific agent to activate it and run them in the background or foreground depending on the task. This way you’re kinda working with a team of developers who follow your instructions.

AI coding best practices

• Learn to review all the code that the agents write. It’s crucial for production systems, where you're working with a team and go beyond “vibe-coding” to actually serving users. Optimise your setup for effective code review.
• Break your feature into the smallest tasks and execute them one by one, step by step, instead of trying to execute a massive project at once. Learn to write clean PRD files with acceptance criteria. Attach requirement files instead of sending long chat messages.
• Separate planning and execution. No need to waste tokens and your cognitive power to realise that AI got your instructions wrong, or they were incomplete. Always ask agent to show a plan first to approve it and get the best output in fewer steps.
• Run multiple chats to solve multiple isolated problems in parallel. Learn to multitask, as each prompt takes time to execute and requires you to switch your focus.
• Use Ralph Technique to solve complex problems and build complex token‑intensive projects. Run multiple AI agents for hours to implement the whole app or to make big refactorings [e.g. this project costs $260 in tokens].
• When iterating on UI, ask to design multiple options. This is a powerful way to build the best interfaces. Use Google AI Studio for prototyping interface layouts (it’s free).
• Send screenshots and images to the agents to provide UI references, explain concepts, or highlight issues. It’s just like you screenshare to your agent. Videos work too. Beware it costs tokens.
• Be mindful of tokens used. Learn to optimise for pricing. It will not just save you money, but also enforce you to use the best workflows, which consume fewer tokens.
• Install code review bots in GitHub: Codex, CodeRabbit, Vercel Bot.
• Keep it secure: as agents have access to your system, bad things can happen when there are vulnerabilities in agents or, for example, VS Code extensions. Consider running your coding environment on a virtual machine. For remote environments, sandboxes are emerging to run untrusted code safely.

Agent coding tools

General purpose agents:

• Cursor (standalone IDE)
• GitHub Copilot (VS Code plugin)
• Open Code (open-source, Terminal app + VS Code plugin)

Locked to a specific LLM:

• Claude Code (Terminal app + VS Code plugin)
• Codex (CLI tool + VS Code plugin)

You can use different models for general purpose agents. Some models work better than others on specific tasks; you should find out which ones work best for you. For example, Gemini works great for front-end code. Claude Opus is great for backend code and now in general is considered the state of the art.

Note: you can use your Copilot and Codex subscriptions for Open Code (but not Claude Code sub).

Pricing (AI costs money)

All AI agents are paid. Basic plans start from $20/mo. Copilot starts at $10/mo.

If you produce lots of code (e.g. building a startup), you’ll likely hit limits soon and will need to upgrade to common $100/mo or $200/mo plans.

Learn to craft the best contexts and optimise for tokens.

Learn AI by building

Engineering has most definitely changed. Now it is the best time to learn the new ways of building, as it's still early. You'll need to unlearn your previous workflows, and it will be painful. But it is well worth it for your productivity and future career.

The only way to learn how to code with AI is to actually build something with AI.

If you can't use AI on your work project, then build a side project. Go crazy and try all the features to understand which ones are most useful. Test different models. Run 10 agents in parallel. Write skills and craft context to see how it changes the output.

Connect with other engineers building with AI and create something cool together.

Just like with any skill: you need to learn and practice, and it takes time!

PS Learn to write well. Markdown is the new JavaScript.