Before it had a name

About a year ago, a few of us put together a rough small on-device agent at work — built in a week, as a proof of concept. I wrote about it not long after, and there was excitement around what it could do, so we kept working on it. At first, feedback was that it was a neat trick, but not the revolution that agents had promised, so it was a lower priority, for a while.

That didn’t last long.

The on-device agent kept coming up, kept drawing interest, and over the few months we had made real progress — not a polished product, but it was enough to get permission to find out whether the idea actually held up. So we kept working at it, in the cracks between everything else we were supposed to be doing, and set out to turn the held-together-with-tape proof of concept into something real.

None of us had built anything like it. We needed it to run on multiple platforms, so we went looking for the best cross-platform option with the fewest moving parts for the front-end, and landed on Flutter — which none of us had ever touched. The backend we wrote in Go. We even had a terminal client written in Bubble Tea. On paper, that’s a stack and a problem we had no business shipping, especially in our spare time.

We shipped it in a little over six months.

What made it possible wasn’t heroics — it was the agent tools. “Learn a new framework and build something with it” used to be a quarter of ramp-up before you wrote anything that mattered. With an agent riding along, learning Flutter was building in Flutter — I’d ask how something worked, get a working example against our actual code, and keep moving. The ramp and the work collapsed into one action.

What we were able to deliver wasn’t a demo. It was a real client, backed by a real service, that could actually do things for you — not “look, it can talk,” but “it can go do the thing you asked.”

The reactions told the whole story. Some people lit up — “wow,” “I had no idea we could do this.” Others squinted: “…why do I need this again?” Both were correct. It was genuinely impressive and not yet obviously necessary — the exact spot some new capabilities sit in right before they become obvious.

But I’d stopped wondering whether this was sustainable. Watching a tiny team, working in their spare time, ship a cross-platform agent that did real work — in a stack we hadn’t known six months earlier — I was sure. Agents that can actually act on your behalf weren’t a someday thing.

Postscript, from later: about five months after this, an open-source project called OpenClaw showed up with more or less the same idea, out in the open — and within a couple of months the whole internet had found it. We’d been quietly building toward the same idea for a year, without a name for it. Watching it land felt like confirmation: we had the right idea all along.