2025-01-12-谷歌-2025年Agents与基础应用白皮书（英文版）_42页_4mb

• Definition: Generative AI agents are autonomous programs that extend beyond standalone models by using tools to access external information, plan actions, and execute tasks independently. For example, an agent can retrieve real-time data or interact with APIs to perform user-driven tasks.

• Core Components: Agents consist of a language model as the central decision maker, tools for interacting with the external world (e.g., APIs, databases), and an orchestration layer for reasoning, planning, and managing multi-turn interactions.

• Agents vs. Models: Agents differ from models by incorporating tools for real-time data access, maintaining session history, and enabling self-directed action execution, whereas models rely on limited training data and lack proactive capabilities.

• Cognitive Architectures: Agents use orchestration layers with reasoning techniques like ReAct (for reasoning and acting), Chain-of-Thought, and Tree-of-Thoughts to guide decision-making and task execution iteratively.

• Tool Types: Three primary tool categories exist:

  • Extensions: Enable agents to execute API calls directly (e.g., for integrated functionalities like database updates or weather fetches).
  • Functions: Allow cue-based client-side execution, delegating API calls while providing fine-grained control, useful for security or timing constraints.
  • Data Stores: Provide dynamic data access through vector embeddings, supporting applications like retrieval-augmented generation for up-to-date information.

• Enhancing Performance: Techniques like in-context learning, retrieval-based learning, and fine-tuning help models adapt to specific tasks, improving tool selection and nuanced reasoning without retraining.

• Implementation: A quick start example with LangChain demonstrates agent creation using tools like Google Search or Place API in Python, achieving multi-step queries through chained orchestration.

• Production Applications: Vertex AI agents offer managed platforms for building scalable applications, integrating user interfaces, debugging tools, and analytics for real-world deployment, such as mixed-expert agents for complex tasks.

• Summary: Agents combine models and external tools to solve complex problems, with the future evolving toward agent chaining and sophisticated reasoning, requiring an iterative approach for customization and scalability in various domains.

To craft summaries, combine key insights from agent components, tools, cognitive architectures, implementation examples, and production applications, ensuring an overview of Generative AI agent capabilities and advancements.