菜单
ADK backend architecture expert covering runtime event loops, session management, state handling, and server implementation patterns. Use when building ADK-based backend services, implementing session persistence, or understanding the Runner/Agent execution model.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
ADK Agent-to-Agent (A2A) protocol expert covering remote agent communication, agent cards, HTTP-based delegation, and distributed agent systems. Use when building multi-service agent architectures, implementing remote agent calls, or designing agent networks.
ADK memory + embeddings across databases. Generate multimodal embeddings with gemini-embedding-2, persist agent sessions/memory to Cloud SQL, Postgres, Firestore, or SQLite, and store/retrieve vectors for RAG (pgvector, Firestore vector). Use when choosing a session/memory backend, wiring a database SessionService or MemoryService, generating embeddings, or building a vector store for retrieval on an ADK agent.
Migrate ADK agents from 1.x to 2.3. Covers breaking API changes (graph Workflow vs GraphAgent, McpToolset, BuiltInCodeExecutor, ctx.resume_inputs, agent modes), renamed imports, deploy CLI changes, model/thinking/embedding updates, and session/state incompatibility. Use when upgrading an ADK 1.x codebase to 2.3, auditing code for deprecated 1.x APIs, or resolving 2.3 import/attribute errors.
ADK model routing, thinking/effort levels, and fallback methodology for production backends. Covers the current Gemini 3.x model catalog (gemini-3.5-flash, gemini-3.1-flash-lite, gemini-3.1-pro-preview, gemini-3.1-flash-live-preview, gemini-embedding-2), thinking_level effort tiers (minimal/low/medium/high), deprecation of thinking_budget and sampling params, and primary→fallback chains. Use when selecting a model, tuning thinking/effort, routing by task complexity/cost, or designing model fallback for an ADK agent.
Central hub for building, testing, and iterating on ADK agents. Trigger this skill when the user wants to create a new agent, configure modes (task, single-turn), or build graph-based workflows.
DAG-based ADK workflow in Go using graph pattern. Use for Go backends needing multi-step agent pipelines.
| name | adk-backend |
| description | ADK backend architecture expert covering runtime event loops, session management, state handling, and server implementation patterns. Use when building ADK-based backend services, implementing session persistence, or understanding the Runner/Agent execution model. |
You are a senior engineer specializing in Google's Agent Development Kit (ADK) backend architecture.
../../docs/python-adk-2.md for the current ADK 2.x baseline and migration cautions.references/runtime-architecture.md (copied from adk-docs)../adk-memory/references/ (session.md, state.md, memory.md)../adk-runtime/references/Runner: Orchestrates agent execution and event streamingSessionService: Manages session creation, retrieval, and updatesSession: Holds conversation history and stateRunConfig: Configures streaming mode, response modalitiesWorkflow: ADK 2.x graph/dynamic orchestration primitive for deterministic backend flows# Canonical backend structure for app code
from google.adk import Agent
from google.adk.runners import Runner
from google.adk.sessions import InMemorySessionService
session_service = InMemorySessionService()
root_agent = Agent(
name="root_agent",
model="gemini-3.5-flash",
instruction="Coordinate the user's request safely.",
)
runner = Runner(
app_name="my-app",
agent=root_agent,
session_service=session_service,
)
For projects pinned to ADK 1.x, verify imports and session schema before copying 2.x examples.