AI

My Journey to Agent-First Development With Air

Some background

I was an active AI user from the moment ChatGPT was first released, and I really liked its UX. Chatting is a familiar activity – I chat with my family, friends, and coworkers every day – so the learning curve was almost non-existent. On top of that, engaging with chatbots was already familiar as well, since they have been a hallmark of customer support interactions since well before ChatGPT came onto the scene. Needless to say, AI became a part of my daily routine almost straight away.

However, I care a lot about the code I push – maybe too much. That’s why I never used AI for coding at work or in production. Searching for information, yes. A one-day pet project, for sure. But coding at work? Oh, no, “never”!

The release of Claude Code didn’t change much. While it was indeed better at coding, I was never a VIM user, so interacting with it was tediously inconvenient (a terminal-based TUI in 2025, really?!). I was only really able to use it for one-day projects.

To make the leap, I needed an AI tool that would let me smoothly switch between code exploration and agent results, as well as allow me to review and comment on those results without changing the app. This would save me time spent copy-pasting tons of text to find where changes were made. I also needed a smoother UX for editing prompts, as the only approach available at the time was writing endless Markdown documents for each ad-hoc task in a separate code-editing application. There had to be a better way.

Air

Full disclosure, I’m a developer with the Air team, so I started to use JetBrains’ agentic development environment from day one for the sake of dogfooding – long before the Air Public Preview began this March. As before, I only used it for pet projects initially. But this time, something was different: I wasn’t as frustrated when the agent failed to do something. I was able to read the agent’s output without exhaustion. I was able to fix small things quickly without switching applications and or needing to adjust to a new view. I could even use my mouse for text editing (wow!) and the shortcuts stored in my muscle memory from mass editor and chat apps.

I started to have the same feeling I had with ChatGPT – Air didn’t require me to learn anything new: All I had to do was chat, use known interaction patterns, and wait for the results. Neither the AI chat in every application since 2022, nor Cursor, nor terminal-based agents, nor GitHub or Telegram bots had ever given me that feeling before.

So, I started to use Air for work, beginning with small-scope tasks that I could quickly fix by hand if needed, working in parallel with IntelliJ IDEA. It noticeably affected my productivity; I was able to work on several tasks at the same time, switching between them during the day and delegating some work to the agents. Interacting with agents inside a familiar environment removed any potential friction I might have experienced if I had been working with them in the terminal. Inside Air, there was no feeling of exhaustion or disorientation – everything just came naturally.

After a while, I noticed that I only needed to use IntelliJ IDEA for specialized tasks like using a debugger if the agent wasn’t able to find a bug, or doing some non-trivial interactive rebasing. Air had everything else I could possibly need. I still find agents to be extremely bad at code design, architecture, and following project patterns, but I can quickly do those things myself in Air’s editor and ask the agent to finish the job.

Gemini

The final turning point was when I needed to implement a Gemini CLI for Air. We had already integrated Claude and Codex by then, and the time for the third agent had come. This sounded exactly like a task for an agent – do something similar to already existing code. No architecture design, no deep thinking – just follow an existing example. So I gave it a try.

It went much less smoothly than I expected. It took me several days, many code reviews, dozens of comments, and a 140-message-long chat exchange, but I ended up with a working, production-ready solution. This might sound like a lot of effort, but while the agent was coding, I was able to work on other things – fix smaller issues, answer Slack messages, and attend meetings. 

In the end, Air completed the task much faster than I could have solo, and I didn’t need to use any other tools alongside it. No terminal, no IntelliJ IDEA, nothing. It was just like it used to be before the advent of AI fragmented developers’ workflows: I only needed a single tool once again, and in Air, I finally had it!

The next gamechanger

I can’t wait until we can bring Air to mobile devices, at least as a web app. This will fill the remote development hole that’s still present – prompting me to write a follow-up post while I make myself a coffee and leave the programming to my computer (or the cloud). Oh, what a brave new world we live in!

So, it turned out that I needed to experience AI in an environment I already felt well at home in before I could fully appreciate its benefit. Anyway, thank you for reading this to the end. Now go try Air – it’s free to download and the best way to see what I mean for yourself.