> ## Content Index > Fetch the complete content index at: https://www.aigl.blog/llms.txt > Use this file to discover other available public pages before exploring further. # Agents - URL: https://www.aigl.blog/agents/ - Published: 2026-09-19T09:08:40.000Z - Updated: 2026-09-19T09:08:40.000Z - Description: A whitepaper from Google explaining how generative AI agents combine a model, tools and an orchestration layer, covering Extensions, Functions and Data Stores, reasoning frameworks such as ReAct, and implementation with LangChain and Vertex AI. - Author: Jakub Szarmach - Tags: Whitepaper, Agentic AI, Technical Controls, Enterprise AI Adoption, AI Security, #aigl-library ## ⚑ Quick Summary Published by Google and dated September 2024, this whitepaper defines a generative AI agent as "an application that attempts to achieve a goal by observing the world and acting upon it using the tools that it has at its disposal". Its purpose is to explain the foundational building blocks of agents and how to implement them. It sets out three essential components β€” the model, the tools and the orchestration layer β€” which together form a "cognitive architecture". Tools are divided into three types that Google models can use: Extensions, which bridge agents to APIs with agent-side execution; Functions, where the model returns a function name and arguments that the client-side application executes; and Data Stores, vector databases used for retrieval augmented generation (RAG). Reasoning frameworks named are ReAct, Chain-of-Thought and Tree-of-Thoughts, illustrated with a step-by-step ReAct loop of question, thought, action, action input, observation and final answer. The paper closes with targeted learning approaches (in-context learning, retrieval-based in-context learning and fine-tuning based learning), a LangChain and LangGraph quick start using the gemini-1.5-flash-001 model, and production applications on Vertex AI, concluding that agent chaining toward a "mixture of agent experts" is a likely direction. ## 🧩 What’s Covered The whitepaper moves from definitions to implementation across nine parts. - **Introduction and definition:** frames agents as programs that extend a model's capabilities with reasoning, logic and access to external information, using a database retrieval tool or API calls as examples. - **Core components:** describes the model as the centralized decision maker, tools as the link to external data and services, and the orchestration layer as the cyclical process of information intake, internal reasoning and next action. - **Agents vs. models:** a comparison chart covering knowledge limits, session history and multi-turn inference, native tool implementation and native logic layers such as CoT, ReAct or LangChain. - **Cognitive architectures:** uses a chef in a busy kitchen analogy and a ReAct sequence listing question, thought, action, action input, observation (repeated N-times) and final answer; also covers ReAct, Chain-of-Thought and Tree-of-Thoughts. - **Tools:** covers Extensions with a Code Interpreter sample, Functions with a travel concierge use case and a JSON function\_call payload, and Data Stores using vector embeddings, SCaNN matching and a five-step RAG request lifecycle. - **Tools recap:** a table comparing Extensions, Function Calling and Data Stores by execution side (agent-side versus client-side) and by use case, including security, authentication and timing constraints. - **Targeted learning:** three approaches β€” in-context learning, retrieval-based in-context learning and fine-tuning based learning β€” each illustrated with a cooking analogy. - **Implementation:** a LangChain and LangGraph agent using SerpAPI and Google Places tools on a multi-stage query, and a section on Vertex AI managed production agents with testing, evaluation and debugging tooling plus a sample end-to-end architecture. ## πŸ’‘ Why it matters? For teams building or reviewing agent systems, the paper supplies a shared vocabulary and a set of design trade-offs. The distinction between agent-side Extensions and client-side Functions is presented as a way to keep API calls out of the agent when security, authentication, timing or human-in-the-loop review require it, and the Data Store section explains the retrieval loop behind RAG. The Vertex AI section names testing, evaluation, measuring agent performance and debugging as part of production deployment. Readers assessing agents for governance purposes get a description of the architecture and its control points, without accompanying risk, assurance or accountability material. ## ❓ What’s Missing The paper is vendor-centred: it describes Google models, Gemini, Vertex AI and related tooling, and states that Extensions, Functions and Data Stores are "the three primary tool types that Google models are able to interact with", so the implementation guidance is not neutral across platforms. Governance, compliance, risk management, safety evaluation and monitoring are absent, as are security threat models for tool use. Costs, latency and failure modes appear only in passing. Reasoning frameworks and tool capabilities are described as they stood "at the time of this publication", so the technical detail dates quickly, and no regulatory or jurisdictional scope is set. ## πŸ‘₯ Best For Engineering and platform teams designing generative AI agent architectures, including developers choosing between Extensions and Function Calling, architects planning RAG-backed agents over structured or unstructured data, and technical reviewers who need a baseline description of agent components before assessing risk, security or governance requirements. ## πŸ“„ Source Details The document is a 42-page English whitepaper titled *Agents*, dated 2 September 2024 and carrying the running header "Agents N September 2024" on each page. Authors are Julia Wiesinger, Patrick Marlow and Vladimir Vuskovic; reviewers and contributors include Evan Huang, Emily Xue, Antonio Gulli and Anant Nawalgaria, with Antonio Gulli, Anant Nawalgaria and Grace Mollison listed as curators and editors. No series, reference number or URL for the document itself is printed. The full text of all 42 pages was available for this review; the endnotes list external sources.