AI agents are so good that some days I don’t code anymore: instead, I manage agents. Given a feature, I discuss the requirements with the agent, discuss possible solutions, decide on a design and ask the agent for an implementation.
I then test and review it. Sometimes it’s great; other times we made the wrong design decision and I start the process again.
I’m still on top of the code, but this doesn’t feel like coding. I’m not thinking about naming variables or off-by-one errors, although I do sometimes think about cache invalidation once in a while. Mostly I’m 1) thinking about the problem, 2) how solving it helps the business and 3) how to describe it in such a way that I will recognize the solution once it’s done. That last point feels much more like managing than coding, especially when you have five agents working simultaneously on different branches.
Some colleagues are riding the AI wave through IDEs and plugins. I find that plain vanilla Vim is more than enough, but I love AI plugins inside GitHub, Linear, and Slack. Asking OpenHands to fix pre-commit issues directly on GitHub feels like heaven.
As agents gain autonomy, the natural place to talk to them are collaboration tools like Slack, GitHub and Linear. There they can easily talk to each other too: agents writing issues, creating PRs, assigning PRs for reviews, everything done transparently through collaboration tools were humans stay on top.
To some engineers this is a bitter lesson, they chose to be engineers precisely because they don’t want to manage. AI inside one IDE will only get you so far though, so you start a second IDE and a third. By then you are already managing, and you might as well do it from collaboration tools. To them I don’t have many words of advice, except to try it.
Next time a cool feature is suggested in a slack conversation, don’t open an IDE, instead @ an agent. It takes another set of skills, but I know you’ll manage it.
