Exploring AI Agents and User Management: Making Cloudflare's Tools Work for You

Meet Lizzie, a developer advocate at Cloudflare, and Nick, a developer experience engineer at Work OS, as they explore the future of automation and authorization in AI-powered agents acting on our behalf.

  • 1. Lizzie is a developer advocate at Cloudflare and Nick is a developer experience engineer at work OS.
  • 2. They make AI demos, including agents and MCP (Multi-Cloud Project) servers that act on their behalf.
  • 3. Examples of such agents include one for auto-voting in the NBA finals and booking tennis courts in San Francisco.
  • 4. The challenge lies in controlling what these agents are allowed to do.
  • 5. While there are many developer-centric tools for creating agents, there is a need for more user-friendly options that non-developers can use.
  • 6. Work OS focuses on authorization and user management for agents.
  • 7. The aim is to provide the same level of credentials and authorization for agents as for user-facing projects.
  • 8. OAuth is no longer just for humans; it's for agents acting on our behalf.
  • 9. Agents also need memory and persistent storage, which Cloudflare provides through its various services like Compute Cloud Workers, AI model hosting, vector database, SQL database, and Durable Objec
  • 10. Other Cloudflare services that can be useful for building agents include Video Streaming, Image Optimization, and Cloudflare workers with bindings for interacting with web apps, websites, and agen
  • 11. Cloudflare's Durable Objects offer fast storage and can be spun up per user for faster retrieval and storage.
  • 12. Lizzie and Nick demonstrate a simple MCP server built using Cloudflare and work OS, which can be easily deployed and run locally.
  • 13. The demo allows the agent to act on behalf of the user, in this case, ordering a shirt from MCP shop and saving the order details in Key Value storage.
  • 14. The agent also has access to user information like name, email address, favorite song, and roles/permissions through the JWT (JSON Web Token).
  • 15. Durable Objects can store data directly on the MCP server for unique user-specific data.
  • 16. With fine-grained authorization, agents could be authorized per line change, tool change, or network connection in the future.
  • 17. An audit trail is essential to keep track of agent interactions, failures, and end results.
  • 18. It's important to treat users as deputies with access to tools that can be used and misused.
  • 19. The code for the demo is available on GitHub, along with a QR code for easy access.
  • 20. Users can order their own MCP t-shirts using the workflow demonstrated in the presentation or by visiting mcp.shop.
  • 21. Additional tools like "pretty please" can be added to customize agents as needed.
  • 22. The presenters encourage the audience to build and share their own MCP servers with Cloudflare.
  • 23. Fine-grained authorization, audit trails, and user trust are crucial aspects of working with agents in the future.
  • 24. As more tasks are delegated to agents, maintaining control and oversight will become increasingly important.

Source: AI Engineer via YouTube

❓ What do you think? What are your thoughts on the ideas shared in this video? Feel free to share your thoughts in the comments!