agent-sdk（Agent SDK 搜索）

Claude Agent SDK - Comprehensive Skill

Overview

The Claude Agent SDK is Anthropic's official framework for building production-ready AI agents with Claude. It provides a high-level abstraction over the Messages API, handling the agent loop, tool orchestration, context management, and extended-thinking patterns automatically.

When to Use the Agent SDK

Use the Agent SDK when:

• Building autonomous agents that need to use tools iteratively
• Implementing agentic workflows with verification and iteration
• Creating subagent hierarchies for complex task decomposition
• Integrating MCP (Model Context Protocol) servers for standardized tools
• Need production patterns like hooks, permissions, and context management

Don't use when:

• Simple single-turn API calls suffice (use Messages API directly)
• No tool use required (standard chat)
• Custom agent loop logic needed (SDK loop is opinionated)

Language Support

• Python: claude-agents package (recommended for most use cases)
• TypeScript: @anthropics/agent-sdk package (Node.js environments)

Both implementations share the same core concepts and API patterns.

Quick Start

Installation

Python:

pip install claude-agents

TypeScript:

npm install @anthropics/agent-sdk

Authentication

Set your API key as an environment variable:

export ANTHROPIC_API_KEY="your-api-key-here"

Or pass it explicitly in code:

from claude_agents import Agent

agent = Agent(api_key="your-api-key-here")

Basic Agent Creation

Python Example:

from claude_agents import Agent

# Create agent with default settings
agent = Agent(
    model="claude-sonnet-4-5-20250929",
    system="You are a helpful assistant focused on accuracy."
)

# Run simple task
result = agent.run("What is 2+2?")
print(result.response)

TypeScript Example:

import { Agent } from "@anthropics/agent-sdk";

const agent = new Agent({
  model: "claude-sonnet-4-5-20250929",
  system: "You are a helpful assistant focused on accuracy.",
});

const result = await agent.run("What is 2+2?");
console.log(result.response);

First Tool Example

Tools extend agent capabilities with external functions:

Python:

from claude_agents import Agent
from claude_agents.tools import Tool

# Define custom tool
def get_weather(location: str) -> dict:
    """Get weather for a location."""
    return {"location": location, "temp": 72, "condition": "sunny"}

weather_tool = Tool(
    name="get_weather",
    description="Get current weather for a location",
    input_schema={
        "type": "object",
        "properties": {
            "location": {"type": "string", "description": "City name"}
        },
        "required": ["location"]
    },
    function=get_weather
)

# Agent with custom tool
agent = Agent(
    model="claude-sonnet-4-5-20250929",
    tools=[weather_tool]
)

result = agent.run("What's the weather in San Francisco?")
print(result.response)

Core Concepts Reference

The Agent Loop

The SDK manages a complete agent loop automatically:

1. Input Processing: User message + system prompt + tools
2. Model Invocation: Claude generates response (text or tool calls)
3. Tool Execution: SDK executes requested tools, handles results
4. Iteration: Results fed back to model, continues until completion
5. Output: Final response with full conversation history

Key Properties:

• Automatic iteration until task completion or max turns
• Built-in error handling and retry logic
• Context management with automatic compaction options
• Token budget tracking and optimization

Context Management

The SDK manages conversation context automatically:

Context Components:

• System prompt (persistent instructions)
• Conversation history (user messages + assistant responses)
• Tool definitions (available capabilities)
• Tool results (execution outputs)

Subagents for Context Isolation:

# Spawn subagent with isolated context
with agent.subagent(
    system="You are a code reviewer focused on security.",
    tools=[security_scan_tool]
) as reviewer:
    review = reviewer.run("Review this code for vulnerabilities: ...")

# Subagent context doesn't pollute parent

Context Compaction: When approaching token limits, the SDK can automatically summarize earlier conversation turns while preserving critical information.

Tools System

Tools are the agent's interface to external capabilities.

Built-in Tools:

• bash: Execute shell commands
• read_file: Read file contents
• write_file: Write data to files
• edit_file: Modify existing files
• glob: File pattern matching
• grep: Content search

Custom Tool Schema:

Tool(
    name="tool_name",              # Unique identifier
    description="What it does",    # Clear capability description
    input_schema={                 # JSON Schema for parameters
        "type": "object",
        "properties": {...},
        "required": [...]
    },
    function=callable              # Python function or async function
)

MCP Integration: The SDK can use Model Context Protocol (MCP) servers as tool providers:

from claude_agents import Agent
from claude_agents.mcp import MCPClient

# Connect to MCP server
mcp_client = MCPClient("npx", ["-y", "@modelcontextprotocol/server-filesystem"])

agent = Agent(
    model="claude-sonnet-4-5-20250929",
    mcp_clients=[mcp_client]
)

Permissions System

Control which tools agents can access:

Allowed Tools (Whitelist):

agent = Agent(
    model="claude-sonnet-4-5-20250929",
    tools=[tool1, tool2, tool3, tool4],
    allowed_tools=["tool1", "tool2"]  # Only these can be used
)

Disallowed Tools (Blacklist):

agent = Agent(
    model="claude-sonnet-4-5-20250929",
    tools=[tool1, tool2, tool3],
    disallowed_tools=["tool3"]  # All except tool3
)

Permission Modes:

• "strict": Agent MUST get permission before tool use (via hooks)
• "permissive": Agent can use allowed tools freely (default)

Hooks System

Hooks provide lifecycle event interception for observability, validation, and control.

Available Hooks:

• PreToolUseHook: Before tool execution (validation, logging, blocking)
• PostToolUseHook: After tool execution (logging, result modification)
• PreSubagentStartHook: Before subagent spawns (context setup)
• PostSubagentStopHook: After subagent completes (result processing)

Basic Hook Example:

from claude_agents.hooks import PreToolUseHook

class LoggingHook(PreToolUseHook):
    async def execute(self, context):
        print(f"Tool: {context.tool_name}")
        print(f"Args: {context.tool_input}")
        return context  # Allow execution

agent = Agent(
    model="claude-sonnet-4-5-20250929",
    hooks=[LoggingHook()]
)

Blocking Tool Use:

class ValidationHook(PreToolUseHook):
    async def execute(self, context):
        if context.tool_name == "bash" and "rm -rf" in context.tool_input.get("command", ""):
            raise PermissionError("Destructive command blocked")
        return context

Common Patterns

File Operations

from claude_agents import Agent

agent = Agent(
    model="claude-sonnet-4-5-20250929",
    allowed_tools=["read_file", "write_file", "glob"]
)

result = agent.run(
    "Read all Python files in ./src and create a summary in summary.md"
)

Code Execution

agent = Agent(
    model="claude-sonnet-4-5-20250929",
    allowed_tools=["bash"]
)

result = agent.run(
    "Run the test suite and analyze any failures"
)

Agentic Search (Gather-Act Pattern)

# Agent automatically gathers information iteratively
search_agent = Agent(
    model="claude-sonnet-4-5-20250929",
    tools=[web_search_tool, read_url_tool]
)

result = search_agent.run(
    "Research the latest developments in quantum computing and summarize key papers"
)

Subagent Delegation

main_agent = Agent(model="claude-sonnet-4-5-20250929")

# Delegate specialized task to subagent
with main_agent.subagent(
    system="You are an expert data analyzer.",
    tools=[analyze_csv_tool, plot_tool]
) as analyzer:
    analysis = analyzer.run("Analyze sales_data.csv and create visualizations")

# Results available to main agent
main_agent.run(f"Based on this analysis: {analysis.response}, what actions should we take?")

Error Handling

from claude_agents import Agent, AgentError

agent = Agent(model="claude-sonnet-4-5-20250929")

try:
    result = agent.run("Your task here", max_turns=10)
except AgentError as e:
    print(f"Agent failed: {e}")
    print(f"Turns completed: {e.turns_completed}")
    print(f"Last message: {e.last_message}")

Verification Pattern

# Agent can self-verify results
agent = Agent(
    model="claude-sonnet-4-5-20250929",
    tools=[calculator_tool, verify_tool]
)

result = agent.run(
    "Calculate the compound interest for $10000 at 5% for 10 years. "
    "Verify your calculation by computing it a second way."
)

Navigation Guide

When to Read Supporting Files

reference.md - Read when you need:

• Deep understanding of agent loop internals
• Complete API reference for all SDK features
• Detailed tool schema specifications
• Permission and security configuration options
• Comprehensive hooks reference with all event types
• Skills system implementation details

examples.md - Read when you need:

• Working code examples for specific patterns
• Tool implementation templates
• Hook implementation examples
• Advanced patterns (subagents, verification, error recovery)
• Integration examples with existing systems

patterns.md - Read when you need:

• Production-ready architectural patterns
• Agent loop optimization strategies (Gather, Act, Verify, Iterate)
• Context management best practices
• Tool design principles
• Security patterns and anti-patterns
• Performance optimization techniques

drift-detection.md - Read when you need:

• Understanding how this skill stays current
• Implementing drift detection for other skills
• Update workflow and validation processes
• Self-validation mechanisms

Integration with Amplihack

The Agent SDK skill integrates with the Amplihack framework:

Creating Amplihack Agents with SDK:

# In .claude/agents/amplihack/specialized/my_agent.md
# Use Agent SDK patterns for tool-using agents
from claude_agents import Agent
from claude_agents.tools import Tool

def create_specialized_agent():
    return Agent(
        model="claude-sonnet-4-5-20250929",
        system="<agent_role_from_md>",
        tools=[...],  # Custom tools for this agent
        hooks=[...]   # Logging, validation hooks
    )

Using MCP Servers in Amplihack:

# Integrate MCP tools into Amplihack workflow
from claude_agents.mcp import MCPClient

mcp_client = MCPClient("npx", ["-y", "@modelcontextprotocol/server-github"])

agent = Agent(
    model="claude-sonnet-4-5-20250929",
    mcp_clients=[mcp_client]
)

# Agent can now use GitHub MCP tools
result = agent.run("Create a GitHub issue for the bug we just found")

Hooks for Amplihack Observability:

# Log all agent actions to Amplihack runtime logs
class AmplihackLoggingHook(PreToolUseHook):
    async def execute(self, context):
        log_to_amplihack_runtime(
            session_id=get_current_session(),
            tool=context.tool_name,
            input=context.tool_input
        )
        return context

Quick Reference

Essential Commands

# Basic agent
agent = Agent(model="claude-sonnet-4-5-20250929")
result = agent.run("task")

# With tools
agent = Agent(model="...", tools=[tool1, tool2])

# With permissions
agent = Agent(model="...", allowed_tools=["tool1"])

# With hooks
agent = Agent(model="...", hooks=[LogHook()])

# Subagent
with agent.subagent(system="...") as sub:
    result = sub.run("subtask")

# MCP integration
from claude_agents.mcp import MCPClient
mcp = MCPClient("npx", ["-y", "mcp-server-name"])
agent = Agent(model="...", mcp_clients=[mcp])

Common Tool Patterns

# File operations
tools=["read_file", "write_file", "glob"]

# Code execution
tools=["bash"]

# All built-in
tools=["bash", "read_file", "write_file", "edit_file", "glob", "grep"]

Token Budget Recommendations

• Simple tasks: 4K-8K tokens
• Complex tasks: 16K-32K tokens
• Research/analysis: 64K-128K tokens
• Maximum context: 200K tokens (model dependent)

Next Steps

1. Start Simple: Create a basic agent with built-in tools
2. Add Custom Tools: Implement tools for your specific domain
3. Add Hooks: Implement logging and validation
4. Use Subagents: Delegate specialized tasks
5. Integrate MCP: Use standardized tool servers
6. Optimize: Tune context, permissions, and verification patterns

For complete API details, see reference.md. For working code, see examples.md. For production patterns, see patterns.md.