The Right Ship

It didn't take long. The thing was genuinely good at code, startlingly good. But ask it to reason carefully through a problem with real stakes, and something was missing. It could write a function faster than any of us, yet it couldn't always hold a line of thought together. We came away with a conviction that stayed with us for years: these models were powerful, but power and reasoning were not the same thing.

So we watched, closely. And the pace was relentless. Reasoning improved. Then tool calls let the models act instead of just talk. Then came retrieval. RAG let them reach outside their own training and ground themselves in real documents. Each of these arrived faster than the last, and each one moved the ceiling higher than we'd expected it to move.

Naturally, we built. We wired up our own MCP servers and our own retrieval pipelines, stitching together the pieces to make the models do useful, specialised work. For a long stretch, that was where our energy went.

And the whole time, quietly, there was a feeling we didn't want to name: vanity. There is something faintly absurd about a small team trying to out-engineer OpenAI and Anthropic on precisely the layer they obsess over, with orders of magnitude more talent and capital than we'll ever have. We kept building, but the question never fully went away.

Then the IDEs arrived. Continue, Windsurf, Cursor, Antigravity, tools that started to feel less like developer toys and more like products that could genuinely go mainstream. We saw it coming and tried to get ahead of it. We forked VSCode and stripped its interface down for people who had never opened a terminal in their lives. We even pitched it to a few law firms here in Switzerland. The response taught us something more valuable than a signed contract would have: we were too early. The idea was right; the timing wasn't.

Through all of it, in the background, two tools never left our own machines: Claude Code and Codex. We used them every single day. And because we used them, we knew, viscerally and not theoretically, how much more powerful they were than any plain chat window. The gap was not subtle. It was the difference between talking to a model and working alongside one.

That daily use also made us nervous about our own plans. We could watch software being commoditised in real time. Building a startup around yet another web app, another dashboard, another thin wrapper, was starting to look like building sandcastles ahead of the tide. The thing we thought we were supposed to make was quietly becoming the thing nobody would pay for.

What did not get commoditised was what people kept coming back to. Claude Desktop. The Codex app. Those were the surfaces that stuck. People began to sense, sincerely, that something powerful was happening inside them: not a gadget, not a demo to show a friend, but something real and durable.

By around December 2025, the shift was impossible to argue with. Agentic AI had reached a kind of maturity. The tools stopped being clever party tricks. They simply worked, reliably enough, for long enough, on hard enough problems, that you could put real weight on them and they wouldn't buckle.

That's when our role finally came into focus. Our contribution was never going to be a better model. That race was already being won by others. Our contribution would be to help people board the best ships. Claude Desktop. The Codex app. The vessels actually capable of carrying serious work across open water.

We believed then, and still believe now, that the endgame is total computer use. Not chat. Not autocomplete. A system that can operate a computer the way a capable, trusted colleague would: opening the files, running the tools, doing the work end to end. Everything before that is a step toward it.

And these apps are beginning to show exactly that. The early signs of complete computer use are already visible inside them. What was a thesis is becoming something you can watch happen on your own screen.

So that is what we do now. We put this new generation of tool, mature and no longer a gadget, directly into our clients' hands. We help them adapt it to their work, customise it to how they actually think, and use it to the very best of its capabilities. The machine is remarkable out of the box; the real value appears when it starts to carry your judgment, your standards, your way of doing things.

And we do it alongside the best people in their fields, the ones whose expertise is worth encoding in the first place. Together we're learning how far these LLM-operated tools can really go, and pushing that frontier a little further every week.

We were early once, and we were wrong about a few things along the way. But we never stopped using the tools we believed in. That conviction is what brought us here, and it's why we're confident about where this is heading.