Hermes Agent 进阶实战

mcp:
  servers:
    - name: filesystem
      command: npx
      args: ["-y", "@modelcontextprotocol/server-filesystem", "/home/user/workspace"]
      env: {}

    - name: github
      command: npx
      args: ["-y", "@modelcontextprotocol/server-github"]
      env:
        GITHUB_PERSONAL_ACCESS_TOKEN: "{GITHUB_TOKEN}"

    - name: puppeteer
      command: npx
      args: ["-y", "@modelcontextprotocol/server-puppeteer"]
      env: {}

    - name: memory
      command: npx
      args: ["-y", "@modelcontextprotocol/server-memory"]
      env: {}

  # 工具调用设置
  tool_timeout: 30  # 秒
  max_retries: 2

在对话中使用 MCP 工具

自定义 MCP 服务器

创建一个自己的 MCP 服务器：

# my_mcp_server.py
from mcp.server import MCPServer
from mcp.types import Tool, TextContent

server = MCPServer("my-tools")

@server.tool()
def get_weather(city: str) -> TextContent:
    """Get current weather for a city."""
    # 接入真实天气 API
    return TextContent(text=f"Weather in {city}: 22°C, sunny")

@server.tool()
def search_jobs(query: str) -> list[dict]:
    """Search for job listings."""
    # 接入招聘 API
    return [{"title": "Senior Dev", "company": "Acme", "url": "..."}]

if __name__ == "__main__":
    server.run()

多模型路由策略

不同任务用不同模型，节省成本的同时保证质量。

配置模型路由

models:
  default: anthropic/claude-sonnet-4-5
  fallback: openai/gpt-4o

  routing:
    # 复杂推理用 Opus
    "complex-reasoning": anthropic/claude-opus-4-6
    # 快速响应用 Haiku
    "fast-response": anthropic/claude-haiku-3
    # 代码任务用 Sonnet（性价比最高）
    "code-generation": anthropic/claude-sonnet-4-5
    # 长文本分析用 Opus
    "deep-analysis": anthropic/claude-opus-4-6
    # 简单问答用 Haiku
    "simple-qa": anthropic/claude-haiku-3

  # 根据 token 数量自动选择
  auto_route_by_length: true
  auto_route_threshold: 32000  # 超过这个长度自动切 Opus

手动指定模型

成本控制策略

Agent Teams 协作

Hermes 支持多 Agent 协作，一个任务拆分给多个 Agent 并行处理。

创建 Agent Team

# ~/.hermes/teams/research.yaml
name: research-team
description: "Multi-agent research workflow"

agents:
  - name: researcher
    model: anthropic/claude-sonnet-4-5
    role: "Researcher - searches and gathers information"
    tools: [web_search, web_extract, memory]

  - name: analyst
    model: anthropic/claude-opus-4-6
    role: "Analyst - analyzes data and draws conclusions"
    tools: [execute_code, memory]

  - name: writer
    model: anthropic/claude-haiku-3
    role: "Writer - formats and writes the final report"
    tools: [file_write, memory]

workflow:
  # researcher 先搜集
  step1:
    agent: researcher
    task: "Research {query} and save key findings to memory"
  # analyst 分析
  step2:
    agent: analyst
    depends_on: [step1]
    task: "Analyze research findings from memory, identify patterns"
  # writer 输出
  step3:
    agent: writer
    depends_on: [step2]
    task: "Format analysis into a markdown report"

触发 Team 任务

上下文压缩与优化

当对话变长时，Hermes 会自动压缩上下文，但也可以手动优化。

配置上下文压缩

context:
  # 自动压缩阈值（token 数）
  compression_threshold: 80000

  # 压缩保留策略
  compression_strategy: "hybrid"  # hybrid | recent | summary

  # hybrid 模式：保留最近 + 重要摘要
  hybrid:
    recent_ratio: 0.6      # 60% 保留最近
    summary_ratio: 0.3    # 30% 保留摘要
    discard_ratio: 0.1    # 10% 丢弃低价值

  # 强制压缩（立即压缩当前上下文）
  /hermes:compress

  # 查看当前上下文大小
  /hermes:context-size

上下文优化技巧

• 长任务分段执行，每段独立压缩
• 重要信息用 @memory 保存，不依赖上下文
• 代码任务避免在同一条消息中混杂多个文件
• 使用 /hermes:clear 重置对话（保留记忆）

Cron 高级用法

依赖外部脚本的 Cron

创建 ~/.hermes/scripts/check_revenue.py：

#!/usr/bin/env python3
"""Check yesterday's revenue metrics from Stripe"""
import requests
import json
from datetime import datetime, timedelta

yesterday = (datetime.now() - timedelta(days=1)).strftime("%Y-%m-%d")
api_key = "sk_live_xxx"

response = requests.get(
    "https://api.stripe.com/v1/charges",
    params={"created[gte]": yesterday, "limit": 100},
    auth=(api_key, "")
)

charges = response.json().get("data", [])
total = sum(c["amount"] for c in charges) / 100

print(f"Date: {yesterday}")
print(f"Total Revenue: ${total:.2f}")
print(f"Transactions: {len(charges)}")
print(f"Avg Order: ${total/len(charges):.2f}" if charges else "No transactions")

失败重试与告警

cron:
  default_retry: 3
  retry_delay: 300  # 5分钟

  on_failure:
    action: telegram_notify
    message: "❌ Cron job '{job_name}' failed: {error}. Retry {retry_count}/3 in 5min."

  on_success:
    action: telegram_notify
    message: "✅ Cron job '{job_name}' completed successfully"

Cron 链式任务

流式响应与工具并行

启用流式输出

# ~/.hermes/config.yaml
display:
  streaming: true        # 流式显示 agent 思考过程
  streaming_speed: 5      # 字符/帧
  show_tool_calls: true  # 显示工具调用

# Telegram 中流式响应
telegram:
  streaming: true
  # Agent 思考过程实时显示在消息中

工具并行执行

Hermes 默认并行执行独立的工具调用（不等一个完成再调用下一个）：

多 Profile 与场景切换

Hermes 支持多 Profile，不同场景使用不同配置，互不干扰。

创建 Profile

Profile 特定配置

# ~/.hermes/profiles/work/config.yaml
models:
  default: anthropic/claude-sonnet-4-5

tools:
  enabled: [terminal, web, file, github, pagerduty]
  disabled: []

telegram:
  bot_token: "work-bot-token"
  admin_ids: [123456]

cron:
  default_deliver: telegram

# ~/.hermes/profiles/my-project/config.yaml
models:
  default: anthropic/claude-opus-4-6

tools:
  enabled: [terminal, file, github]
  disabled: []

github:
  token: "project-token"

生产级监控

日志与审计

健康检查

Alerting 配置

alerts:
  # Cron 连续失败告警
  cron_failure:
    threshold: 3        # 连续失败 3 次
    action: telegram_notify
    message: "⚠️ Cron job '{job_name}' failed {count} times consecutively"

  # 模型响应超时
  model_timeout:
    threshold: 60        # 超过 60 秒无响应
    action: telegram_notify
    message: "⚠️ Model response timeout ({duration}s)"

  # API Key 余额不足
  api_balance:
    threshold: 10        # 余额低于 $10
    action: telegram_notify
    message: "⚠️ API balance low: ${balance}"