A few years ago I pasted small snippets from a chat window into my editor. Today an agent edits my project directly, runs it, takes screenshots of the result and deploys it. This is how that happened, stage by stage: what I did at each step, and what the industry was building at the same time.
Each section starts with a mind map. Hover a node for a summary and what was special about it; click or tap it for how it works, news and numbers.
Before AI: autocomplete and searching for answers
Before any of this, the editor could list the methods of an object or expand a snippet, and everything else came from documentation and search. Nothing wrote logic for you.
Stage 1: AI chat writes a snippet
My first AI code came from a chat window. I described a small problem, got a snippet back and pasted it into my project. It was good for a function or a regular expression, not for a feature.
My journey
This is where most developers first met AI coding. ChatGPT arrived in November 2022 and made it normal to ask for code in plain words. The chat knew nothing about your project, so the answers were small and general.
Stage 2: Copilot in the editor: hints and explanations
Then the AI moved into my editor. GitHub Copilot suggested the next lines, explained code and proposed solutions. It did not implement the feature for me: I still reviewed every suggestion and wrote the real change myself.
My journey
Copilot actually launched in 2021, a year before ChatGPT, but many developers (me included) started with chat and met Copilot second. At this stage the AI is a pair programmer that whispers: it helps you type and understand, and you stay the one who builds.
Stage 3: Chat gets my files and context
Chat got much better. I could give it whole files and the context of the project, and it generated or edited exactly the code I wanted. But the result lived in the chat window, so I copied it back into the IDE by hand.
My journey
Stronger models and much larger context windows changed what chat could do: from a snippet to a whole file rewritten correctly. The remaining problem was the gap between the chat and the project.
Stage 4: Agents change the code directly
With Cursor, Claude Code and Codex the AI works inside the project. It adds, edits and generates code directly, runs it and fixes what fails. Most days I no longer read every line: I describe the result I want and check that it works.
My journey
An agent is a model in a loop with tools. It edits files, runs commands, reads the output and tries again. Some agents work beside you in the editor or terminal, others take a task and come back with a pull request.
Stage 4, continued: What agents now do on their own
Each model release is stronger than the last, and the tools around the model keep growing. My agent runs commands, takes screenshots to look at its own work, uses MCP servers and connects to my servers to deploy and read logs.
My journey
The model is only half of an agent. The other half is what it is allowed to touch: the terminal, the browser, outside services and remote machines. Each new connection removes one more thing the developer used to do by hand.
Vibe coding: when you stop reading the code
I can now build a feature without opening a single file. That is exactly what people call vibe coding, and it is worth being honest about when that is fine and when it is not.
My journey
Not reading the code is a real choice with real costs. For a prototype or a personal tool it is a gift. For anything with users, money or private data, someone still has to understand what was built.
What this means if you are learning to code
Across all four stages one thing never moved to the AI: deciding what to build and judging whether the result is right. The tools write more of the code, so the skills that remain yours matter more.
At every stage the AI took over one more step I used to do by hand: typing, then writing a file, then copying it into place, then running and checking it. What it never took over is knowing what I want and judging whether I got it.
Want to practise the skills from the last section? Write and run code in the online Python compiler, or test yourself with coding challenges that check your solution automatically.