谷歌云技术指南：打造智能AI代理系统

Table ofcontents Introduction 01 An overview of Google Cloud’s agent ecosystemKey components of every agentThe role of grounding in agentic systemsKey takeaways04091723 How to build AI agents25 A complete toolkit for building AI agentsA step-by-step guide: Defining an LLM agentGovern and scale your agent workforce with Google AgentspaceOther options for building agentsKey takeaways2740434546 Ensuring AI agents are reliable and responsible48 AgentOps: A framework for production-ready agentsBuild responsible and secure AI agents with AgentOpsKey takeaways505456 More from Google’s full AI stack58 Conclusion 59 Resources 60 Introduction The development of AI agents represents a paradigm shift insoftware engineering, enabling startups to automate complexworkflows, create novel user experiences, and solve businessproblems that were previously technically infeasible. Whether you’re validating an idea, building an MVP,or supporting a product in production, this guidewill help across all stages of your project. How to use this guide But moving from a promising prototype to a production-readyagent means solving a new set of challenges. How do youmanage their non-deterministic behavior? How do you verifytheir complex reasoning paths? And, crucially, where do youget started? New to AI agents? Start withSection 1for the core concepts. This technical guide will help answer questions like these.It provides a systematic, operations-driven roadmap fornavigating the new landscape, and is geared to help startupsand developers who are racing to embrace the potentialof agentic systems. Ready to build? Jump toSection 2to create your first agent using ADK. Agent built? Dive intoSection 3to make it safe, stable, and scalable. You’ll learn the foundational concepts of agentic systems,from their core architectural components to the principlesthat ensure reliable and responsible operation in production.And you’ll learn about the full spectrum of tools that makebuilding and using agents on Google Cloud more efficient,from code-first development withAgent Development Kit(ADK) and operational automation with theAgent Starter Pack,to no-code agent creation with Google Agentspace. Want extra support? Use the Gemini Kitto prototype faster, and applyto theGoogle for Startups Cloud Programto receiveexpert guidance and up to $350k USD in cloud credits. The focus of this guide The agentic AI ecosystem offers many tools, libraries,and approaches for building cognitive architectures.There are open-source frameworks from Google likeGenkit andGoogle Cloud’s conversational AI offerings,as well as popular open-source libraries like LangChainand CrewAI. This guide focuses primarily on ADK, sharing conceptsand architectural patterns that allow you to build robust,scalable agents on Google Cloud while retaining the abilityto integrate other preferred tools and libraries. Core conceptsof AI agentsSection 1g The field of agentic AI is evolving rapidly.This section provides foundational knowledgeon AI agents, explaining their core concepts,purpose, and operational mechanics. It alsodetails the relevant tools and services availablewithin Google Cloud. This podcast was created using NotebookLM with theprompt: “As a podcast host, create a conversational andeducational podcast for ‘Startup technical guide: AI agents,’targeting a technical audience of startup founders anddevelopers. The podcast must cover the three main pathsfor using AI agents (build, use, partner), detailing toolslike the Agent Development Kit (ADK) and pre-builtGemini agents. Prefer audio? Listen to the podcast versionof this section, created with NotebookLM. “It should then explain the core components of an agent,including models, tools, orchestration, and runtime. Also,cover how to ensure trust and power through techniqueslike grounding with Retrieval-Augmented Generation (RAG)and leveraging multimodality. Conclude with a summaryof the key takeaways and a clear call to action directinglisteners to Google’s resources.” An overview of GoogleCloud’s agent ecosystem1.1 The agentive workflow is the next frontier.It’s not just about asking a question andgetting an answer. It’s about giving AI acomplex goal—like ‘plan this product launch’or ‘resolve this supply chain disruption’—and having it orchestrate the multi-steptasks needed to achieve it. This willfundamentally change productivity.” Building production-grade AI agents requires more thanselecting a large language model. A complete solutiondemands scalable infrastructure, robust data integrationtooling, and architectural patterns that accommodatediverse technical requirements. Google Cloud supports the comprehensive developmentof agentic systems, whether you’re building your own agents,using pre-built Google Cloud agents, or bringing in partneragents. Underpinned by theModel Context Protocol (MCP)and Agent2Agent(A2A) protocol, this common frameworkis designed for interoperability. This way, regardless