future-agi/workshop-demo

This project is a hands-on demo for a Future AGI workshop. It shows how to take a basic AI assistant and make it smarter using a platform called FutureAGI, which helps you test, evaluate, and improve your AI's underlying instructions. The repo walks through three example assistants: a simple chatbot, a bike rental assistant that can book bikes, and a voice-based hotel receptionist. Each of the three assistants has two versions. The first is a basic version with a prompt that is intentionally just "good but not great." The second version connects to FutureAGI to run simulations, evaluate the AI's responses, and automatically optimize the prompt. This setup lets you directly compare a standard AI assistant against one that has been refined through automated testing and improvement. The audience for this project is developers or technical founders attending a Future AGI workshop who want to learn how to systematically improve their AI agents. For example, if you built a bike rental chatbot and customers complain it misunderstands booking dates, this project demonstrates how to simulate conversations, evaluate where the bot fails, and automatically rewrite its instructions to fix those issues. A notable aspect of the project is that it is designed to work with multiple AI providers. You can use models from OpenAI, Anthropic, or Groq by simply changing a setting in an environment file. The README does not go into detail on exactly how the three different optimizers (MetaPrompt, ProTeGi, and GEPA) work, but it makes clear that each of the three assistants relies on a distinct optimization method provided by FutureAGI.