I Wrote This in 2023. Nobody Cared. Now Everyone's Doing It.

Two years ago, I started exploring how LLMs could move from text completion to real-world orchestration. Actually doing things, not just talking. Way before it was trending.

I built a prototype. In Go. With a custom JavaScript-like scripting language that allowed us to define actions dynamically and safely. The LLM would translate human prompts into these scripts, and the scripts would be executed in a tightly controlled environment. It worked. It was flexible, secure, and honestly, kind of cool.

To help people understand why and how, I wrote an article. But it never saw the light of day. And to be fair, that might've been for the best—the ecosystem wasn't ready, and neither was I. But now that everyone’s doing it, I’m finally talking about it.

But now, suddenly, the AI landscape has caught up. Everyone is trying to build "LLM agents," stitch together autonomous workflows, and use models to act on real systems. All those early concepts? They're mainstream. And I can't help but wonder: what if we had kept going?

This series is my way of reopening that door. Not out of spite (okay, maybe a little), but because the original idea still feels valuable. It had architectural benefits, business potential, and a level of control that's more relevant than ever.

If there is interest, I’ll follow up through what that system was supposed to be, what worked in theory and in practice, and how I’m thinking about building it again—with the benefit of hindsight and far better tools.

If you’ve ever had a good idea too early, or too quietly, you’ll get it. And if you're building one of these systems now, maybe you'll find something useful—or at least familiar—in what comes next. Code and all.

continue