AI in early 2026
The way I do my job has been changing very quickly in the last year or so. In a conversation a few days ago, I remarked that I probably won't remember the details of this time well, a few years from now.
So it seems worthwhile to describe it a little, if only for posterity. I expect most engineers are experiencing something like this, though most are probably more conservative (and some further ahead than me).
Thinking back to about a year ago, in May 2025, the extent of my AI usage was a combination of GitHub Copilot and ChatGPT. I'd been using both for several years at that point. The tab completion model of Copilot was very powerful, a useful speedup that acknowledged the limitations of the models of the time - even small snippets could be wrong, and were easily ignored or improved by supplying more context. I was a huge fan and spent quite some effort getting my employer at the time to purchase it for all the engineers. I suppose over the years I used it, it improved a little, but I never found the newer features like chatting or code reviewing particularly useful.
ChatGPT was obviously very helpful, though the limitations of needing to copy and paste back and forth between an IDE and a webpage made it clunky. Still, it was great for writing one-off scripts or fixing the kinds of esoteric build issues which would have once taken quite a lot of time.
So by May I'd been using these tools for a year and was pretty happy. Around this time Cursor, billed as an AI-native IDE, was becoming quite popular. I tried it a few times, but swapping IDE is a big hassle, and one without enough payoff to justify it - I felt a bit guilty about that, given how popular it was becoming.
But around this time I started experimenting with Claude Code, and this was just a huge upgrade. The CLI based workflows meant I didn't have to give up my preferred IDE, and I was still much too fussy about my code to push stuff without manual tweaks. So from May 2025 I was using Claude heavily, but pretty much everything I'd do would have a lot of my own changes on top.
The next step was in November 2025, when Opus 4.5 became available. This was when the capabilities of the model hit a point where it became my default first port of call for pretty much any task - I recall making a post in my work slack in late November specifically extolling its virtues.
The final change so far has been the move over to Conductor. I mentioned above that I loved Claude Code because it let me keep my old IDE, but by the time of Opus 4.6 and 4.7 I found that I didn't need it for 80% of changes - I'd specify a change, and first read the code as a PR, and generally make changes through chat rather than manually. I think part of this was improvements in the models and getting the right memory and customization settings to handle my tastes, but a lot was simply me becoming less fussy. As I've worked with agents more I start to see the minor details of code style as less important.
And this is where Conductor fits in - as a kind of control center for managing lots of changes at once, it lets me dip into my IDE when I need to, but otherwise focus on managing agents working on many things in parallel.
So as of May 2026 Conductor is where I spend my day-to-day doing engineering, occasionally opening IntelliJ for a worktree for things where I think the code style is important or there's something I don't understand. But it seems very likely this isn't the final shape software engineering will take, and I'm very curious to see what'll happen next!