Research pulls back curtain on Claude
Anybody who's been playing around with AI this year in any meaningful way will have noticed that the "models" have taken a back seat to the "orchestration layer," the non-AI operational tasks that allow the technology to function in the real world (If this were a fancy restaurant, the AI model would be the Michelin star chef while the orchestration would be all the wait staff keeping the restaurant running.)
The real power of AI models comes when they are harnessed by an elaborate piece of software that lets users interact with the technology, whether it's Codex, Claude Cowork, or OpenClaw. "Only about 1.6% of Claude Code's codebase constitutes AI decision logic, with the remaining 98.4% being operational infrastructure," a little-noticed research paper from April explains. In fact, the AI model itself is sort of cordoned off from anything important. For that, old-fashioned, deterministic computer code handles everything.
These still aren't really consumer products — yet. For that to happen, we need vastly more complex software to handle everything that comes after the AI. Consumers don't want to think about things like security or access control. They want it all to just work.
Of course, you still need a very good AI model to get great results. But as time goes on, the AI models themselves become less important than all the infrastructure being built around them.
