Agent基建三栈：网关·运行时·协议实战评测

graph TB
    subgraph Protocol["协议层 - Agent互操作"]
        A2A["A2A Protocol<br/>Google · Linux Foundation"]
        ACP["ACP Protocol<br/>IBM Research · BeeAI"]
        MCP["MCP Protocol<br/>Anthropic · 广泛生态"]
    end

    subgraph Runtime["运行时层 - Agent执行平台"]
        AC["AWS Bedrock<br/>AgentCore"]
        GE["Google Gemini<br/>Enterprise Agent Platform"]
        MS["Microsoft<br/>Agent365 / Foundry"]
    end

    subgraph Gateway["网关层 - Agent流量管理"]
        AG["agentgateway<br/>Solo.io · Rust"]
        PK["Portkey<br/>多模型路由"]
        LL["LiteLLM<br/>开源统一API"]
        KG["Kong AI Gateway<br/>企业级插件"]
    end

    Runtime -->|"部署目标"| Gateway
    Protocol -->|"通信标准"| Runtime
    Protocol -->|"发现与委派"| Gateway

    style Protocol fill:#e8f5e9
    style Runtime fill:#e3f2fd
    style Gateway fill:#fff3e0

# agentgateway典型配置片段
apiVersion: gateway.agentgateway.io/v1alpha1
kind: AgentGateway
spec:
  listeners:
    - protocol: MCP
      port: 8080
      routes:
        - backend: mcp-server-tools
          authentication:
            type: jwt
            issuer: https://auth.example.com
    - protocol: A2A
      port: 8081
      routes:
        - backend: remote-agents
          tls:
            mode: MUTUAL
  observability:
    tracing: opentelemetry
    metrics: prometheus

# Portkey统一路由示例
from portkey_ai import Portkey

portkey = Portkey(
    api_key="PORTKEY_API_KEY",
    config={
        "strategy": "fallback",
        "targets": [
            {"provider": "anthropic", "model": "claude-sonnet-4-6"},
            {"provider": "openai", "model": "gpt-4.1"},
        ],
        "cache": {"type": "semantic", "ttl": 3600},
        "guardrails": ["pii-filter", "content-safety"]
    }
)

response = portkey.chat.completions.create(
    messages=[{"role": "user", "content": "分析这份合同的风险条款"}]
)

# AgentCore CLI快速部署示例
# 安装CLI
# pip install @aws/agentcore

# 初始化Agent项目
# agentcore init my-agent --runtime python3.12

# 部署到AgentCore Runtime
# agentcore deploy --memory vector --identity my-agent-role

# Antigravity部署配置示例
apiVersion: antigravity.dev/v1
kind: Agent
metadata:
  name: financial-analyst
spec:
  model: gemini-2.5-pro
  tools:
    - mcp://bigquery-analytics
    - a2a://risk-assessment-agent
  memory:
    type: persistent
    backend: cloud-spanner
  identity:
    serviceAccount: [email protected]
  scaling:
    minInstances: 2
    maxInstances: 100

// Agent Card示例
{
  "@context": "https://a2a-protocol.org/v1",
  "name": "risk-assessment-agent",
  "description": "评估金融交易风险等级",
  "url": "https://agents.example.com/risk-assessor",
  "capabilities": {
    "streaming": true,
    "pushNotifications": true
  },
  "authentication": {
    "schemes": ["oauth2", "api-key"]
  },
  "skills": [
    {
      "id": "assess-transaction-risk",
      "name": "交易风险评估",
      "input": {"type": "object", "properties": {"amount": {"type": "number"}}},
      "output": {"type": "object", "properties": {"riskLevel": {"type": "string"}}}
    }
  ]
}

stateDiagram-v2
    [*] --> Submitted: Client发送Task
    Submitted --> Working: Remote Agent接受
    Working --> Working: 流式更新(Artifact)
    Working --> Completed: 成功完成
    Working --> Failed: 执行失败
    Working --> Cancelled: 客户端取消
    Submitted --> Rejected: Remote Agent拒绝
    Completed --> [*]
    Failed --> [*]
    Cancelled --> [*]
    Rejected --> [*]

graph LR
    subgraph Agent侧
        Agent["AI Agent"]
    end

    subgraph 三协议定位
        MCP["MCP<br/>Agent → 工具/数据<br/>Anthropic发起"]
        A2A["A2A<br/>Agent → Agent<br/>Google发起"]
        ACP["ACP<br/>Agent → Agent(轻量)<br/>IBM发起"]
    end

    subgraph 资源侧
        Tools["外部工具<br/>API/数据库"]
        Agents["其他Agent"]
    end

    Agent -->|"调用工具"| MCP
    Agent -->|"委派任务"| A2A
    Agent -->|"轻量协作"| ACP
    MCP --> Tools
    A2A --> Agents
    ACP --> Agents

[业务Agent] → [agentgateway] → [MCP Server集群]
                    ↓
              [A2A发现服务] → [协作Agent]
                    ↓
              [LLM路由(Portkey)] → [多模型后端]
                    ↓
              [可观测性(OTel)] → [Grafana]

[Agent365/Foundry] ─── 治理层 ──→ [策略引擎]
[AgentCore Runtime] ── 执行层 ──→ [沙箱容器]
[Agent Gateway] ────── 网关层 ──→ [MCP/A2A路由]
                                        ↓
                    ┌───────────────────┼───────────────────┐
                    ↓                   ↓                   ↓
              [内部工具MCP]      [外部Agent A2A]      [LLM Provider API]

graph TD
    Start["选择Agent基础设施"] --> Cloud{"已有云厂商偏好?"}
    Cloud -->|"AWS"| AWS["AWS AgentCore<br/>全套托管"]
    Cloud -->|"GCP"| GCP["Gemini Enterprise<br/>Antigravity部署"]
    Cloud -->|"Azure"| Azure["Agent365/Foundry<br/>企业治理"]
    Cloud -->|"无/混合"| Self{"团队规模?"}

    Self -->|"≤5人"| Small["agentgateway自建<br/>Docker Compose<br/>MCP+A2A"]
    Self -->|"5-20人"| Med["agentgateway + K8s<br/>Portkey LLM路由<br/>OTel可观测"]
    Self -->|"20+人"| Large["多集群agentgateway<br/>混合云Runtime<br/>A2A联邦发现"]

    AWS --> NeedA2A{"需要跨Agent协作?"}
    GCP --> NeedA2A
    Azure --> NeedA2A
    NeedA2A -->|"是"| Full["启用A2A协议<br/>agentgateway做网关"]
    NeedA2A -->|"否"| MCPOnly["MCP满足<br/>成本最优"]

    style Start fill:#4CAF50,color:white
    style AWS fill:#FF9800,color:white
    style GCP fill:#2196F3,color:white
    style Azure fill:#9C27B0,color:white