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

That’s a meaningful split — authoring workspace vs execution engine with persistent state and lifecycle controls is exactly the kind of separation that matters. Checkpoints and recoverability are where most “just prompt it” approaches collapse. There’s real architectural overlap with what ORCA does, though the scope is different. ORCA generalizes that pattern: any cognitive task — not just narrative generation — gets decomposed into reusable capabilities with typed contracts, bound to swappable backends (Python, OpenAPI, MCP), and executed through a runtime with DAG scheduling, step-level tracing, and checkpoint/restore. The key difference is that ORCA’s capabilities and skills are domain-agnostic and open. A text.content.summarize capability used inside a story pipeline is the same one used in a legal document workflow or a security audit — same contract, same governance, different binding. Sounds like you’ve solved the vertical problem well for narrative. ORCA is trying to solve the horizontal one. Would genuinely be interested to compare notes when you release. The more systems that take state persistence and execution structure seriously, the better the argument gets for everyone.