Green squares

For a few years, my GitHub graph told a quiet story. A green square here and there, long gray stretches in between. Almost all of it came through work, and the team I led kept a light open-source footprint — so my GitHub commits added up to a handful a year. I’d moved into managing a team a while back, and between that and life, the time to sit down and build something had mostly evaporated. My list of personal projects just kept getting longer and my GitHub graph grayer.

It’s not that the tools weren’t there. I’d had my first real experiences with agentic coding about eighteen months earlier, and at work the assistants kept getting better. But “better” had a ceiling. They were good for a single task, or a small handful — but even then they often lacked the context to know what you actually wanted, or had to be told exactly how to solve a problem before they could solve it. They hallucinated, a lot. They stayed a curiosity. Most of the engineers around me reached for them for research and the occasional bug fix; when it was time to write something real, they closed the tab and wrote it themselves.

The tools aimed at everything else — drafting an email, writing a doc — were rougher still. Often wrong, sometimes not even in the right ballpark, and they sounded like robots. You could always tell when something had been written by one. It had no soul. There were tells — the breathless “Great question!”, the flowery padding, the relentless enthusiasm — it felt a little like the thing took about as much time away as it saved.

But there was still something there. You could stand up a proof of concept in an afternoon, in a language you’d never touched, and actually learn something — and in the early, uncertain phase of a project, that alone was helpful.

Then 2025 began.

The pace picked up, slowly at first. A new capability every quarter or so. Then every couple of months. Then every month. By the end of the year it was every week — I heard more than one senior engineer say, only half-joking, that they’d given up trying to keep up. I knew the feeling.

Then, in November, Antigravity landed alongside Gemini 3, and something shifted. Gemini 3 wasn’t a better autocomplete; it was actually a useful collaborator. It could hold a session long enough to do real work. You still had to dance around the rough edges — context bloat mostly — but it was workable in a way that it hadn’t been before. Reddit was a wall of discussion about Cursor, Claude, Codex and Antigravity. The announcements only got faster.

Watching it all scroll past, I realized that things had gotten helpful, and I was very interested in things that are useful.

The list of projects I’d been meaning to build for years suddenly didn’t feel so out of reach. Antigravity looked genuinely good — the planning mode, the inline comments, the whole shape of it looked like something I’d enjoy working on. So one evening I stopped watching and opened it up. It all came back quickly, and I remembered how much I loved seeing my ideas come to life as I built them.

The graph was about to get a lot greener.