“It’s the Architecture, Stupid” — Why Prompt Engineering Won’t Fix Agents

Glad to hear that — and I mean it. The “models are compute power” framing is exactly right. That’s the core thesis: the model is an execution engine, not the architecture itself. Curious about what you’ve built — especially around reuse and composition of cognitive steps. Most people who arrive at this conclusion independently end up solving similar structural problems (persistence, structured I/O, separation of reasoning from execution) but with very different tradeoffs. If you’re open to sharing, I’d genuinely like to compare notes. The point of publishing ORCA wasn’t to claim novelty on every front — it was to make the architectural argument explicit and give it a concrete, reproducible runtime. The more people building in this direction, the faster we move past the “just prompt harder” era.