View a markdown version of this page

构建人工智能代理 - Amazon Nova
创建代理调用代理

构建人工智能代理

Amazon Nova 模型针对使用 Amazon Nova Act 构建人工智能代理进行了优化。这些模型优化了工具使用能力，提升了多步骤任务推理效果，增强了在复杂代理工作流中维持上下文的能力，并支持远程 MCP 工具。

创建代理

使用 Nova 构建的人工智能代理可协调多次工具调用，在持续交互过程中保持上下文，并在需要时进行纠错调整。扩展思考能力通过对复杂目标开展系统性推理，可优化代理工作流。建议采用 Strands Agents 等规划框架 SDK，使代理系统的规划与执行流程更加稳定可靠。

代理设计模式

使用 Nova 设计代理时：

  • 对于需要规划与验证的复杂多步骤工作流，建议将推理等级设为中级或高级，以便获得最佳效果

  • 启用 auto 工具选择，支持在代理交互过程中灵活选用工具

  • 设计错误处理机制，使代理能够恢复执行并采用调整后的方案重试

  • 保留对话历史，在代理交互全过程中维持上下文信息

  • 针对代理系统处理的非受控内容，部署完善的内容过滤与审核机制。例如，Amazon 提供 Amazon Bedrock 护栏功能，可在多个基础模型、知识库和代理中应用安全防护策略。此类护栏可过滤有害内容、屏蔽违规话题，并对个人身份信息等敏感数据进行脱敏处理。

多工具代理示例

tool_config = { 
    "tools": [ 
        { 
            "toolSpec": { 
                "name": "calculator", 
                "description": "Perform mathematical calculations", 
                "inputSchema": { 
                    "json": { 
                        "type": "object", 
                        "properties": { 
                            "expression": { 
                                "type": "string", 
                                "description": "Mathematical expression to evaluate" 
                            } 
                        }, 
                        "required": ["expression"] 
                    } 
                } 
            } 
        }, 
        { 
            "toolSpec": { 
                "name": "database_query", 
                "description": "Query financial database for historical data", 
                "inputSchema": { 
                    "json": { 
                        "type": "object", 
                        "properties": { 
                            "query": { 
                                "type": "string", 
                                "description": "SQL query to execute" 
                            } 
                        }, 
                        "required": ["query"] 
                    } 
                } 
            } 
        } 
    ] 
} 
 
response = client.converse( 
    modelId=" us.amazon.nova-2-lite-v1:0", 
    messages=[{ 
        "role": "user", 
        "content": [{ 
            "text": "Analyze our Q3 financial performance across all business units, calculate year-over-year growth rates with statistical significance testing, and recommend budget allocation strategies for Q4." 
        }] 
    }], 
    toolConfig=tool_config, 
    inferenceConfig={"maxTokens": 10000, "temperature": 1, “topP”: 0.9}, 
    additionalModelRequestFields={ 
        "reasoningConfig": { 
            "type": "enabled", 
            "maxReasoningEffort": "low" 
        } 
    } 
)

调用代理

代理调用涉及管理对话流程、处理工具调用以及在多次交互之间维护状态。

流式代理响应

通过流式响应实时呈现代理的推理过程与执行操作：

import boto3
response = client.converse_stream( 
    modelId=" us.amazon.nova-2-lite-v1:0", 
    messages=[{ 
        "role": "user", 
        "content": [{ 
            "text": "Design a scalable microservices architecture for an e-commerce platform handling 1M+ daily transactions. Consider data consistency, fault tolerance, performance, security, and cost optimization." 
        }] 
    }], 
    inferenceConfig={"maxTokens": 10000, "temperature": 10}, 
    additionalModelRequestFields={ 
        "reasoningConfig": { 
            "type": "enabled", 
            "maxReasoningEffort": "low" 
        } 
    } 
) 
 
# Process the streaming response 
reasoning_complete = False 
for event in response["stream"]: 
    if "contentBlockDelta" in event: 
        delta = event["contentBlockDelta"]["delta"] 
         
        if "reasoningContent" in delta: 
            reasoning_text = delta["reasoningContent"]["reasoningText"]["text"] 
            print(f"{reasoning_text}", end="", flush=True) 
        elif "text" in delta: 
            if not reasoning_complete: 
                print(f" 
 
Final Architecture Design: 
") 
                reasoning_complete = True 
            print(f"{delta['text']}", end="", flush=True)

代理状态管理

维护对话历史与工具执行结果以保留上下文；以下示例仅演示单轮对话场景，但开发人员可根据工作流需求自行编排整个代理系统。此外，Strands 等 Amazon Web Services 工具可代表开发人员管理代理上下文和工具状态。

messages = []

messages = [] 
 
# Initial user query 
messages.append({ 
    "role": "user", 
    "content": [{"text": user_query}] 
}) 
 
# Get agent response 
response = client.converse( 
    modelId=" us.amazon.nova-2-lite-v1:0", 
    messages=messages, 
    toolConfig=tool_config, 
    inferenceConfig=inf_params 
) 
 
# Add assistant response to history 
messages.append(response["output"]["message"]) 
 
# Process tool calls and add results 
if response["stopReason"] == "tool_use": 
    tool = next( 
        block["toolUse"] 
        for block in response["output"]["message"]["content"] 
        if "toolUse" in block 
    ) 
     
    # Execute tool 
    result = execute_tool(tool["name"], tool["input"]) 
     
    # Add tool result to conversation 
    messages.append({ 
        "role": "user", 
        "content": [{ 
            "toolResult": { 
                "toolUseId": tool["toolUseId"], 
                "content": [{"json": result}], 
                "status": "success" 
            } 
        }] 
    }) 
     
    # Continue conversation 
    response = client.converse( 
        modelId=" us.amazon.nova-2-lite-v1:0", 
        messages=messages, 
        toolConfig=tool_config, 
        inferenceConfig=inf_params 
    )

代理最佳实践

有关代理最佳实践的更多信息，请参阅一般最佳实践。

有关开发对话式人工智能代理的指南，请参阅语音转语音（Amazon Nova 2 Sonic）。