In 2023 I built the first version of an engine where AI models reason together, because one model alone kept getting the hard calls wrong.
By 2023 large language models had gone from research to daily tool, and everyone rushed to bolt them onto existing software. The pattern that broke down fastest was the single assistant asked to handle a long, high-stakes workflow on its own, where one confident wrong step poisoned everything after it. In space, a spacecraft returned a sample from an asteroid, a reminder that the hardest problems are solved in careful, verified stages, not one heroic move.
By 2023 the tools I had waited a decade for were finally here. Language models could read, write, and reason well enough to change how software worked. So like everyone else, I started putting them to work, and I ran straight into the wall that a single model hits on any long, serious task.
One model asked to plan, decide, and execute on its own is confident even when it is wrong, and a single wrong step early poisons every step after it. In a restaurant, in an enterprise, in anything with real stakes, that failure mode is not acceptable. Adding AI to software was easy. Making it trustworthy was the actual problem.
So I built the first version of a different approach: not one assistant, but specialised models that reason against each other, critique, verify, and improve the work before it ever reaches a human decision. It behaved less like a chatbot and more like a small engineering team arguing toward a better answer. That first rough engine is where everything I am building now begins. The question I have been chasing since is simple to say and hard to earn: how do you make software that reasons before a human ever has to decide.