Summary
Wes Roth introduces Ghost AI (ghost.build), a tool that creates isolated, disposable PostgreSQL database copies on demand—solving one of the most underappreciated risks in agentic AI development: giving agents write access to stateful data. The video opens with Roth’s GravellGPT benchmark, a custom physics simulation where models including Claude Code Opus 4.7 and GPT-5.5 write ship control scripts and iteratively improve over 30 attempts and 200 simulation ticks based on collision and fuel feedback. As the project grew, the danger of agents touching a shared database became a real constraint.
The core argument is that while AI agents can safely experiment in branched code—changes are diffable and reversible—databases have always been the exception. A database is live application state: users, orders, game economies, analytics events. Ghost addresses this by letting agents spin up parallel database copies with identical schemas but fully isolated state, enabling A/B experiments across different economy configurations or pricing logic without risking production data. Each attempt can be thrown away like a draft.
The practical walkthrough demonstrates Ghost’s MCP integration, which allows CodeX, Claude Code, Cursor, Windsurf, and other agentic coding tools to create and manage database snapshots directly from the agent context. Roth shows three parallel sub-agents each working on a separate Ghost database instance for his AI Village project—a multi-LLM simulated world with villagers controlled by different models. The broader takeaway positions disposable database infrastructure as a natural next step as agentic workflows move from editing files to managing application state.
📺 Source: Wes Roth · Published May 30, 2026
🏷️ Format: Hands On Build
