Google AI Agent 启动技术指南 (Thomas Kurian)（Z-Library）

# Google AI Agent 启动技术指南 ## 【One-Line Pitch】 A practical, operations-first playbook for startups and developers who want to move AI agents from promising prototypes to reliable, scalable production systems on Google Cloud—covering everything from core concepts and architecture patterns to hands-on tooling with the Agent Development Kit (ADK). ## 【Book Arc】 - **Opening (~0%–9%)**: Introduces the AI agent paradigm shift and maps the Google Cloud agent ecosystem—three paths to adoption: building custom agents (code-first with ADK or app-first with Agentspace), using pre-built Google Cloud agents, and integrating partner agents. Establishes interoperability via MCP and A2A protocols as the foundation. - **Early (~9%–27%)**: Breaks down the anatomy of an agent—model selection (balancing capability, speed, cost), grounding through RAG and its evolution toward GraphRAG and Agentic RAG, orchestration patterns (ReAct loop), and the runtime infrastructure needed for production deployment. - **Early–Middle (~27%–41%)**: Deepens the grounding discussion with concrete Google Cloud implementations—vector databases, Vertex AI Search, and the three-tier memory architecture (long-term knowledge, working memory, transactional memory) mapped to specific data services like Firestore, Memorystore, Cloud SQL, and BigQuery. - **Middle (~41%–55%)**: Transitions to the build phase—introduces the ADK as the sweet spot between low-code speed and full flexibility. Covers the three agent types (LLM-based, workflow agents, custom logic), the ReAct loop implementation, tool design principles, and the MCP standardization layer. - **Late (~55%–end)**: Moves into production concerns—managing data across Google Cloud services, memory distillation as the next frontier, AgentOps for production-grade agent operations, and building responsible AI agents. Concludes with resources for further learning and Google Cloud startup support. ## 【Key Takeaways】 - **Model selection is about trade-offs, not raw power** (Early): Choose the most efficient model for each task—Gemini 2.5 Flash-Lite for high-throughput/low-latency work, Flash for balanced production apps, Pro for complex reasoning. Multi-agent architectures let you mix models dynamically, reserving heavy models for hard reasoning while light models handle routine queries. - **Grounding is what makes agents trustworthy** (Early): RAG moves agents beyond pretrained knowledge by retrieving verifiable facts before generating answers. The progression from basic RAG → GraphRAG (understanding relationships) → Agentic RAG (active, multi-step retrieval) represents increasing sophistication in how agents interact with knowledge. - **Orchestration via ReAct turns LLMs into problem-solvers** (Early): The Reason-Act-Observe loop is the core pattern—the agent forms hypotheses, calls tools, observes results, and iterates. This enables multi-step workflows like refund processing, customer onboarding, and proactive system monitoring that go far beyond single-turn Q&A. - **Memory is a three-tier architecture** (Early): Long-term knowledge (grounding, context, analytics) uses services like Vertex AI Search and BigQuery; working memory (conversation context, caching) needs sub-millisecond access via Memorystore; transactional memory (audit trails, state changes) requires ACID guarantees from Cloud SQL or Cloud Spanner. - **ADK offers three agent types for different execution modes** (Middle): LlmAgent provides flexible, non-deterministic reasoning; workflow agents (Sequential, Parallel, Loop) deliver predictable, deterministic execution; CustomAgent (BaseAgent subclass) gives full control via Python code. Your choice determines the reasoning-vs-control balance. - **Tools are API contracts for models** (Middle): Effective tool design requires descriptive function signatures, precise docstrings (the semantic core), and structured return patterns with status keys. ADK supports everything from simple FunctionTools to Agent-as-a-Tool delegation and RemoteA2aAgent for cross-process communication. - **MCP is the universal adapter for agent ecosystems** (Middle): The Model Context Protocol standardizes how agents connect to external tools and data sources—ADK agents can act as MCP clients (using third-party tools) or wrap their own tools as MCP servers (exposing them to the broader ecosystem). - **Memory distillation is the next frontier** (Late): As conversation histories grow, feeding full context becomes inefficient and costly. LLM-driven memory distillation—asynchronously extracting key facts and preferences into compact, structured long-term memory—enables personalized, continuous experiences without the overhead. Vertex AI Memory Store (preview) offers early implementations. ## 【Reading Tips】 - **Skim the opening ecosystem overview** (~0%–9%): The three-path framework (build/use/partner) is useful context, but you can move quickly if you already know you want to build custom agents—jump ahead to the component breakdown. - **Deep-read the memory architecture section** (Early, ~18%–23%): The three-tier memory model (long-term, working, transactional) is the conceptual backbone for the entire data strategy. Understanding this mapping to specific Google Cloud services will save you significant architectural rework later. - **Focus on the ADK agent types and tool design** (Middle, ~45%–55%): This is the most actionable content for builders. The distinction between LlmAgent, workflow agents, and CustomAgent—and when to use each—directly determines your system's behavior and maintainability. - **Treat the ReAct loop as your mental model**: The Reason-Act-Observe pattern appears repeatedly throughout the book. If you internalize this single concept, the rest of the architecture (tools, orchestration, memory) falls into place naturally. - **Skim the Google Cloud service catalogs**: The book includes extensive service descriptions (Firestore, BigQuery, Cloud SQL, etc.). Read these as reference material—note which services exist and what they're for, but don't try to memorize details until you need them for a specific use case. ## 【Coverage Limits】 Excerpts cover the conceptual foundations, architecture patterns, and ADK tooling in depth, but do not include actual code examples, step-by-step tutorials, or detailed configuration guides. The AgentOps section and responsible AI discussion are mentioned but only briefly sampled in the available material. ##