aws-agentcore-langgraph — aws-代理core-langgraph

AWS 代理Core + LangGraph

Multi-代理 系统s on AWS Bedrock 代理Core with LangGraph orchestration. Source: https://github.com/aws/bedrock-代理core-启动er-工具kit

安装 pip 安装 bedrock-代理core bedrock-代理core-启动er-工具kit langgraph uv 工具 安装 bedrock-代理core-启动er-工具kit # 安装s 代理core 命令行工具

Quick 启动 from langgraph.graph 导入 状态Graph, 启动 from langgraph.graph.message 导入 添加_messages from langgraph.prebuilt 导入 工具Node, 工具s_condition # routing + 工具 execution from bedrock_代理core.运行time 导入 Bedrock代理Core应用 from typing 导入 Annotated from typing_扩展s 导入 TypedDict

class 状态(TypedDict): messages: Annotated[列出, 添加_messages]

构建器 = 状态Graph(状态) 构建器.添加_node("代理", 代理_node) 构建器.添加_node("工具s", 工具Node(工具s)) # prebuilt 工具 executor 构建器.添加_conditional_edges("代理", 工具s_condition) # 路由s to 工具s or END 构建器.添加_edge(启动, "代理") graph = 构建器.compile()

应用 = Bedrock代理Core应用() # Wraps as HTTP 服务 on port 8080 (/invocations, /ping) @应用.entrypoint def invoke(payload, 上下文): 结果 = graph.invoke({"messages": [("user", payload.获取("prompt", ""))]}) return {"结果": 结果["messages"][-1].content} 应用.运行()

命令行工具 Commands Command Purpose 代理core 配置 -e 代理.py --region us-east-1 设置up 代理core 配置 -e 代理.py --region us-east-1 --name my_代理 --non-interactive Scripted 设置up 代理core launch --部署ment-type contAIner 部署 (contAIner mode) 代理core launch --disable-memory 部署 without memory sub系统 代理core dev Hot-reload local dev server 代理core invoke '{"prompt": "Hello"}' Test 代理core destroy 清理up Core Patterns Multi-代理 Orchestration 编排器 delegates to specia列出s (customer 服务, e-commerce, 健康care, financial, etc.) Specia列出s: inline functions or separate 部署ed 代理s; all 分享 会话_id for 上下文 Memory (STM/LTM) from bedrock_代理core.memory 导入 Memory命令行工具ent memory = Memory命令行工具ent() memory.创建_event(会话_id, actor_id, event_type, payload) # Store 事件 = memory.列出_事件(会话_id) # Retrieve (returns 列出)

STM: Turn-by-turn within 会话 | LTM: Facts/decisions across 会话s/代理s ~10s eventual consistency after writes Gateway 工具s python -m bedrock_代理core.gateway.部署 --stack-name my-代理s --region us-east-1

from bedrock_代理core.gateway 导入 Gateway工具命令行工具ent gateway = Gateway工具命令行工具ent() 结果 = gateway.call("工具_name", param1=value1, param2=value2)

Transport: Fallback Mock (local), Local MCP servers, Production Gateway (Lambda/REST/MCP) Auto-配置s BEDROCK_代理CORE_GATEWAY_URL after 部署 Decision Tree Multiple 代理s coordinating? → 编排器 + specia列出s pattern Persistent cross-会话 memory? → 代理Core Memory (not LangGraph 检查points) External APIs/Lambda? → 代理Core Gateway Single 代理, simple? → Quick 启动 above Complex multi-step 记录ic? → 状态Graph + 工具s_condition + 工具Node

Key Concepts 代理Core 运行time: HTTP 服务 on port 8080 (handles /invocations, /ping) 代理Core Memory: Managed cross-会话/cross-代理 memory LangGraph Routing: 工具s_condition for 代理→工具 routing, 工具Node for execution 代理Core Gateway: 转换s APIs/Lambda into MCP 工具s with auth Naming Rules 启动 with letter, only letters/numbers/underscores, 1-48 chars: my_代理 not my-代理 Troubleshooting Issue Fix on-demand throughput isn't supported Use us.anthropic.claude-* inference 性能分析s 模型 use case detAIls not submitted Fill Anthropic form in Bedrock Console Invalid 代理 name Use underscores not hyphens Memory empty after write WAIt ~10s (eventual consistency) ContAIner not reading .env 设置 ENV in Dockerfile, not .env Memory not working after 部署 检查 记录s for "Memory enabled/disabled" 列出_事件 returns empty 检查 actor_id/会话_id match; event['payload'] is a 列出 Gateway "Unknown 工具" Lambda must strip ___ prefix from bedrock代理Core工具Name 平台 mismatch 警告 Normal - CodeBuild handles ARM64 cross-平台 builds References 代理core-命令行工具.md - 命令行工具 commands, 部署ment, lifecycle 代理core-运行time.md - 流ing, a同步, observability 代理core-memory.md - STM/LTM patterns, API reference 代理core-gateway.md - 工具 integration, MCP, Lambda langgraph-patterns.md - 状态Graph de签名, routing reference-architecture-advertising-代理s-use-case.pdf - Example multi-代理 architecture