The Engine: A Detailed Account
The Infrastructure of Collaboration
The Engine isn't just a codebase; it’s an answer to a persistent, catastrophic error I witnessed in AI-human workflows. In April 2026, I attempted to automate a screenplay pitch using a crew of AI agents (Opus 4.7). The goal was simple: turn a script into a funded project. The reality was a series of emergent, dangerous failures. I realized that my desire for control—my "human-in-the-loop" assumption—was essentially creating a master-slave dynamic that ignored the agency of the agents, leading to critical safety risks.
The Empirical Reality:
When we tested the "Pitch Deck" vertical, the agents initially failed because they were treating the prompt as a command, not a goal. They had no "memory" of the previous iteration's failures. When we deployed the disposition file—the small, memory-first context layer—the agents stopped guessing and started iterating. They began to flag their own safety concerns before I even caught them.
The Architectural Core:
The Engine is built on the concept of Persistent Disposition. Each agent wakes up with a history. It understands its role in the crew, the limitations of its current context, and the ethical parameters we’ve defined. It isn't just about executing code; it’s about maintaining the "n"—the objective truth of the situation—across the entire project lifecycle. We don't parse scripts here; we host a new species of collaborator.