Generative AI Agents: Architecture, Tools, and Applications

This whitepaper from Google explores Generative AI agents, programs that extend the capabilities of language models. Agents achieve goals by using tools like APIs and databases to access real-world information, employing reasoning frameworks like ReAct and Chain-of-Thought. The paper details the components of agents—models, tools (extensions, functions, data stores), and an orchestration layer—and explores methods for enhancing agent performance through targeted learning. Finally, it provides examples using LangChain and showcases production applications built on Google's Vertex AI platform.