How We Trained an AI Agent on $2 5M Process

Key Concepts

• Vertical AI Agent: An AI agent specialized for a specific task or industry.
• Sora 2 Ads: A platform or tool used for generating advertisements, likely with AI capabilities.
• Foundational Documents: Pre-existing documents or information used as a basis for AI training or prompt generation.
• Clot: A tool or platform used for research, likely in conjunction with AI prompt generation.
• Strategy Agent: An AI agent responsible for developing and executing strategies.
• Read Foundational Document Tool: A specific tool enabling an AI agent to access and utilize foundational documents.
• Mermaid Graph: A text-based diagramming tool that can generate visual representations of processes or data structures.
• Excali draw: A virtual whiteboard tool for creating diagrams and visual representations.
• Handoffs (Communication Flow): A method of communication between AI agents where control is passed from one agent to another, often requiring user feedback.

Training a Vertical AI Agent on a Real Business Process

This section details the methodology for training a vertical AI agent, specifically focusing on an e-commerce advertising process. The core idea is to leverage a publicly shared, real-world business process to train the AI agent effectively.

The Five-Step Process for Sora 2 Ads

The video transcript highlights a five-step process for generating Sora 2 ads, as demonstrated by an e-commerce brand owner who reportedly generated $2.5 million using this method. This process, which can be set up in approximately one hour, forms the basis for training the AI agent.

Foundational Documents and Agent Access

The e-commerce brand owner also shared foundational documents used for research prior to generating prompts for Sora 2. These documents are crucial for the AI agent's training. The trained strategy agent can access these documents at every stage of the process through a specialized "read foundational document tool."

Extracting and Visualizing the Process

The process was extracted from the video transcript using a straightforward method:

1. Transcript Copying: The entire transcript of the video was copied.
2. AI Prompting for Visualization: The transcript was then pasted into ChatGPT with a prompt to extract the process as a Mermaid graph.
3. Diagram Creation: The generated Mermaid graph was copied and inserted into Excali draw to create a visual representation of the process.

System Training and Agent Design

Following the extraction and visualization, the system was trained based on this identified process. The decision was made to implement three distinct AI agents.

Communication Flows: The Role of Handoffs

For communication between these three agents, a "handoff" mechanism was chosen. This approach is deemed suitable because each step within the described process necessitates close feedback and interaction from the user.

Synthesis and Conclusion

The primary takeaway is the effectiveness of training vertical AI agents by replicating real-world, proven business processes. By leveraging publicly available information, such as video transcripts and shared documents, and employing tools like ChatGPT and Excali draw for process extraction and visualization, a robust training framework can be established. The use of specialized tools for document access and a handoff-based communication flow between agents ensures that the AI can effectively follow and execute the complex, user-feedback-dependent steps of the business process. This approach aims to create AI agents that are not only knowledgeable but also practically applicable to generating tangible business results, as exemplified by the $2.5 million e-commerce ad revenue.