Google Cloud explores AI agents in Web3, focusing on automating blockchain interactions. Two primary agent models exist: agent-controlled, where the agent holds the key and manages funds, and transaction-crafter, where the agent creates transactions for user approval. The agent-controlled model introduces performance, malice, and security risks by giving agents direct key control. The transaction-crafter model prioritizes security by having the user sign all transactions. Google Cloud provides tools like the Agent Development Kit (ADK) for building agents and using Gemini 2.0 Flash. The transaction-crafter model is significantly safer as it requires explicit user approval via their wallet. The document suggests that MCP servers should support both models to enable autonomous agents and secure user interactions. This dual approach will foster innovation while maintaining user protection within Web3 applications. The author emphasizes the importance of user control and the flexibility of the transaction-crafter approach.