claude-sdk-expert（Claude SDK expert 搜索）

Claude SDK Expert Skill

Purpose

This skill provides comprehensive guidance on building autonomous AI agents using the Claude Agent SDK (formerly Claude Code SDK), leveraging computer use capabilities, tool orchestration, and MCP integration for production deployments.

SDK Overview

Claude Agent SDK (2025)

The Claude Agent SDK enables building autonomous agents that can interact with computers, write files, run commands, and iterate on their work.

Evolution: Renamed from "Claude Code SDK" to reflect broader capabilities beyond coding.

Core Philosophy: Give Claude a computer to unlock agent effectiveness beyond chat-based interactions.

Key Capabilities

1. Computer Use

Revolutionary Feature: Claude can control a computer environment to complete tasks.

What This Enables:

• File system operations (read, write, edit)
• Terminal command execution
• Iterative debugging and refinement
• Multi-step autonomous workflows
• Real-world task completion

Use Cases:

• Finance agents analyzing portfolios
• Personal assistants booking travel
• Customer support handling complex requests
• Development agents building software
• Research agents gathering and analyzing data

2. Built-in Tools

File Operations:

• Read - Read file contents
• Write - Create or overwrite files
• Edit - Make targeted edits to existing files

Command Execution:

• Bash - Run shell commands and scripts

Search & Discovery:

• Grep - Search file contents with regex
• Glob - Find files by pattern

Web Access:

• WebFetch - Retrieve and analyze web pages
• WebSearch - Search the internet for information

All tools are production-tested and optimized for agent use.

3. MCP Integration

Model Context Protocol Support: Define custom tools via MCP servers.

Benefits:

• Standardized tool interface
• Reusable across different agents
• Community ecosystem of MCP servers
• Enterprise data source connectivity

Example MCP Servers:

• GitHub, Slack, Google Drive
• PostgreSQL, MongoDB
• Stripe, Salesforce
• Custom internal APIs

Architecture Patterns

Pattern 1: Autonomous Task Completion

Scenario: Agent completes multi-step task without human intervention

Flow:

User Request
    ↓
Claude analyzes task
    ↓
Breaks into subtasks
    ↓
Executes via tools (Read, Bash, Write, etc.)
    ↓
Iterates on failures
    ↓
Returns result

Example:

from anthropic import Anthropic

client = Anthropic()

response = client.messages.create(
    model="claude-sonnet-4-5",
    max_tokens=4096,
    tools=[
        {"type": "computer_use"},
        {"type": "bash"},
        {"type": "file_operations"}
    ],
    messages=[{
        "role": "user",
        "content": "Analyze the last 30 days of sales data and create a summary report"
    }]
)

# Claude autonomously:
# 1. Reads sales data files
# 2. Runs analysis scripts
# 3. Generates report
# 4. Saves to file

Pattern 2: Human-in-the-Loop Approval

Scenario: Agent proposes actions, waits for approval before executing

Flow:

Task → Plan → Show to Human → Approve? → Execute → Result
                                ↓ No
                            Revise Plan

Implementation:

# Step 1: Generate plan
plan_response = client.messages.create(
    model="claude-sonnet-4-5",
    messages=[{
        "role": "user",
        "content": "Create a plan to refactor the authentication system"
    }]
)

# Step 2: Human reviews plan
if human_approves(plan_response.content):
    # Step 3: Execute with tools
    execution_response = client.messages.create(
        model="claude-sonnet-4-5",
        tools=all_tools,
        messages=[{
            "role": "user",
            "content": f"Execute this plan: {plan_response.content}"
        }]
    )

Pattern 3: Iterative Refinement

Scenario: Agent iterates on work based on feedback/errors

Flow:

Attempt 1 → Error → Analyze → Attempt 2 → Error → Analyze → Attempt 3 → Success

Built-in: Claude SDK naturally supports this through computer use - agents can see command outputs and adjust.

Tool Design Best Practices

Custom Tool Creation

Good Tool Design:

# Clear, focused tool
{
    "name": "get_customer_orders",
    "description": "Retrieve all orders for a specific customer ID",
    "input_schema": {
        "type": "object",
        "properties": {
            "customer_id": {
                "type": "string",
                "description": "The unique customer identifier"
            },
            "since_date": {
                "type": "string",
                "description": "ISO date to filter orders from (optional)"
            }
        },
        "required": ["customer_id"]
    }
}

Poor Tool Design:

# Too broad, unclear purpose
{
    "name": "do_customer_stuff",
    "description": "Does various things with customers",
    "input_schema": {
        "type": "object",
        "properties": {
            "action": {"type": "string"},
            "data": {"type": "object"}
        }
    }
}

Tool Selection Principles

DO: ✅ Provide tools relevant to the task ✅ Use clear, descriptive names ✅ Write detailed descriptions (Claude reads these!) ✅ Define strict input schemas ✅ Implement error handling in tools ✅ Return structured, parseable outputs

DON'T: ❌ Give agents tools they don't need (increases confusion) ❌ Use ambiguous names like "handler" or "processor" ❌ Skip input validation ❌ Return raw error messages without context ❌ Make tools with side effects unclear

MCP Integration Patterns

Connecting MCP Servers

# Define MCP server connection
mcp_config = {
    "servers": {
        "github": {
            "command": "npx",
            "args": ["-y", "@modelcontextprotocol/server-github"],
            "env": {
                "GITHUB_TOKEN": os.getenv("GITHUB_TOKEN")
            }
        },
        "postgres": {
            "command": "docker",
            "args": ["run", "mcp-postgres-server"],
            "env": {
                "DATABASE_URL": os.getenv("DATABASE_URL")
            }
        }
    }
}

# Claude automatically discovers tools from MCP servers
response = client.messages.create(
    model="claude-sonnet-4-5",
    mcp_servers=mcp_config,
    messages=[{
        "role": "user",
        "content": "Find all GitHub issues assigned to me and update the project database"
    }]
)
# Claude uses both github and postgres MCP tools

Custom MCP Server

# Create custom MCP server for internal API
from mcp import Server, Tool

server = Server("internal-crm")

@server.tool()
def get_customer_data(customer_id: str):
    """Retrieve customer information from internal CRM"""
    return crm_api.get_customer(customer_id)

@server.tool()
def update_customer_notes(customer_id: str, notes: str):
    """Add notes to customer record"""
    return crm_api.update(customer_id, {"notes": notes})

# Deploy and connect to Claude

Production Best Practices

1. Streaming for UX

Why: Show user progress in real-time, build trust in agent actions

with client.messages.stream(
    model="claude-sonnet-4-5",
    max_tokens=4096,
    tools=tools,
    messages=messages
) as stream:
    for event in stream:
        if event.type == "content_block_delta":
            print(event.delta.text, end="", flush=True)
        elif event.type == "tool_use":
            print(f"\nUsing tool: {event.name}")

2. Error Handling

Robust Error Management:

try:
    response = client.messages.create(
        model="claude-sonnet-4-5",
        tools=tools,
        messages=messages
    )
except anthropic.APIError as e:
    # Handle API errors
    log_error(f"API Error: {e}")
    return fallback_response()
except anthropic.RateLimitError:
    # Handle rate limits
    time.sleep(60)
    retry()
except Exception as e:
    # Handle tool execution errors
    log_error(f"Tool Error: {e}")
    return safe_error_message()

3. Cost Optimization

Strategies:

• Use Claude Haiku for simple tasks, Sonnet for complex reasoning
• Implement caching for repetitive contexts
• Batch similar requests when possible
• Limit max_tokens appropriately
• Monitor token usage via callbacks

# Use appropriate model for task
simple_task_response = client.messages.create(
    model="claude-haiku-4",  # Cheaper, faster
    messages=[{"role": "user", "content": "Format this JSON"}]
)

complex_task_response = client.messages.create(
    model="claude-sonnet-4-5",  # More capable
    messages=[{"role": "user", "content": "Analyze architectural trade-offs"}]
)

4. Security

Critical Security Measures:

Tool Permissions:

# Restrict file access
safe_file_tools = {
    "read": {
        "allowed_paths": ["/data/public"],
        "denied_paths": ["/etc", "/secrets"]
    },
    "write": {
        "allowed_paths": ["/output"],
        "denied_paths": ["/"]
    }
}

Input Sanitization:

def sanitize_bash_command(cmd: str) -> str:
    """Prevent dangerous commands"""
    dangerous = ["rm -rf", ":(){ :|:& };:", "dd if="]
    for danger in dangerous:
        if danger in cmd:
            raise SecurityError(f"Dangerous command blocked: {danger}")
    return cmd

Audit Logging:

def log_agent_action(action: dict):
    """Track all agent actions for security audit"""
    audit_log.write({
        "timestamp": datetime.now(),
        "tool": action["tool_name"],
        "input": action["input"],
        "user": action["user_id"],
        "result": action["result"]
    })

Performance Optimization

Parallel Tool Calls

Claude can use multiple tools simultaneously when appropriate:

# Claude automatically parallelizes when possible
response = client.messages.create(
    model="claude-sonnet-4-5",
    tools=[weather_api, stock_api, news_api],
    messages=[{
        "role": "user",
        "content": "Give me weather, stock prices, and news for San Francisco"
    }]
)
# Claude calls all 3 APIs in parallel

Caching Strategies

# Cache system prompts and large contexts
response = client.messages.create(
    model="claude-sonnet-4-5",
    system=[{
        "type": "text",
        "text": large_system_prompt,
        "cache_control": {"type": "ephemeral"}
    }],
    messages=messages
)
# System prompt cached for ~5 minutes

Testing Agents

Unit Testing Tools

def test_customer_lookup_tool():
    """Test individual tool behavior"""
    result = get_customer_orders("CUST123")
    assert result["customer_id"] == "CUST123"
    assert isinstance(result["orders"], list)

Integration Testing

def test_agent_workflow():
    """Test agent using multiple tools"""
    response = client.messages.create(
        model="claude-sonnet-4-5",
        tools=[tool1, tool2, tool3],
        messages=[{
            "role": "user",
            "content": "Process order #12345"
        }]
    )

    # Verify expected tool usage
    tool_calls = extract_tool_calls(response)
    assert "verify_order" in tool_calls
    assert "process_payment" in tool_calls

Evaluation Framework

# Use Claude's built-in evaluation
from anthropic import Anthropic

eval_client = Anthropic()

eval_results = eval_client.evaluate(
    agent=my_agent,
    test_cases=[
        {"input": "...", "expected_output": "..."},
        # More test cases
    ],
    metrics=["accuracy", "latency", "tool_efficiency"]
)

Common Patterns

Pattern: Multi-Step Research

async def research_agent(query: str):
    """Agent researches topic using multiple sources"""
    response = await client.messages.create(
        model="claude-sonnet-4-5",
        tools=[web_search, web_fetch, summarize],
        messages=[{
            "role": "user",
            "content": f"Research '{query}' and provide comprehensive summary"
        }]
    )
    # Claude: searches → fetches articles → summarizes → synthesizes
    return response.content

Pattern: Code Generation & Testing

def code_agent(requirements: str):
    """Agent writes and tests code"""
    response = client.messages.create(
        model="claude-sonnet-4-5",
        tools=[write_file, bash, read_file],
        messages=[{
            "role": "user",
            "content": f"Write and test code for: {requirements}"
        }]
    )
    # Claude: writes code → saves file → runs tests → fixes errors → retries
    return response.content

Pattern: Data Pipeline

def data_pipeline_agent(source: str, destination: str):
    """Agent ETL pipeline"""
    response = client.messages.create(
        model="claude-sonnet-4-5",
        tools=[read_file, bash, postgres_insert],
        messages=[{
            "role": "user",
            "content": f"Extract data from {source}, transform it, and load to {destination}"
        }]
    )
    # Claude orchestrates full ETL
    return response.content

Model Selection

Claude Sonnet 4.5 (claude-sonnet-4-5)

Best For:

• Complex reasoning and analysis
• Multi-step autonomous tasks
• Code generation and debugging
• Research and synthesis
• High-stakes decisions

Characteristics:

• Highest capability
• Best for computer use
• More expensive
• Slower than Haiku

Claude Haiku 4 (claude-haiku-4)

Best For:

• Simple, well-defined tasks
• Format conversions
• Quick classifications
• High-throughput scenarios
• Cost-sensitive applications

Characteristics:

• Fast responses
• Lower cost
• Good for structured tasks
• Limited complex reasoning

Integration Examples

With FastAPI

from fastapi import FastAPI
from anthropic import Anthropic

app = FastAPI()
client = Anthropic()

@app.post("/agent/task")
async def run_agent_task(task: dict):
    response = client.messages.create(
        model="claude-sonnet-4-5",
        tools=load_tools_for_task(task),
        messages=[{
            "role": "user",
            "content": task["description"]
        }]
    )
    return {"result": response.content}

With LangChain (via LangChain-Anthropic)

from langchain_anthropic import ChatAnthropic
from langchain.agents import initialize_agent

llm = ChatAnthropic(model="claude-sonnet-4-5")
agent = initialize_agent(
    tools=[tool1, tool2],
    llm=llm,
    agent_type="structured-chat-zero-shot-react-description"
)
result = agent.run("Complete this task")

Monitoring & Observability

Key Metrics

• Tool Call Success Rate - % of tool invocations that succeed
• Task Completion Rate - % of user requests fully resolved
• Average Iterations - How many tool calls per task
• Latency - Time to complete requests
• Token Usage - Input + output tokens per request
• Error Rate - % of requests with errors

Logging Best Practices

import logging

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger("claude-agent")

def run_agent_with_logging(task):
    logger.info(f"Starting task: {task}")

    response = client.messages.create(
        model="claude-sonnet-4-5",
        tools=tools,
        messages=[{"role": "user", "content": task}]
    )

    logger.info(f"Tools used: {extract_tools(response)}")
    logger.info(f"Token usage: {response.usage}")

    return response

Decision Framework

Use Claude SDK when:

• Building on Anthropic models (Claude family)
• Need computer use capabilities (file, bash, iteration)
• Want production-ready agent framework
• Require MCP integration for data sources
• Building autonomous task completion agents

Consider alternatives when:

• Committed to OpenAI ecosystem (use AgentKit)
• Need visual agent builder (use AgentKit)
• Require complex state machines (use LangGraph)
• Want full OSS control (use AutoGen/LangGraph)

Resources

Official Documentation:

• Agent SDK Docs: https://docs.claude.com/en/api/agent-sdk
• Computer Use Guide: https://docs.anthropic.com/en/docs/agents/computer-use
• MCP Integration: https://modelcontextprotocol.io

GitHub:

• Python SDK: https://github.com/anthropics/claude-agent-sdk-python
• TypeScript SDK: https://github.com/anthropics/claude-agent-sdk-typescript

Final Principles

1. Computer Use is Game-Changing - Leverage file/bash capabilities fully
2. Tools are First-Class - Design tools as carefully as prompts
3. MCP for Data - Use MCP servers for enterprise data connectivity
4. Stream for UX - Real-time feedback builds user trust
5. Security Always - Validate inputs, restrict permissions, audit actions
6. Right Model for Task - Haiku for simple, Sonnet for complex

*This skill ensures you build powerful, autonomous agents using Claude's cutting-edge capabilities in 2025.*