majiayu000 Claude SDK Expert Skill
Purpose
Build autonomous AI agents using Claude Agent SDK, leveraging computer use, tool orchestration, and MCP integration for production deployments.
SDK Overview
Claude Agent SDK (2025)
Enables building autonomous agents that control computers, write files, run commands, and iterate on work.
Core Philosophy: Give Claude a computer to unlock agent effectiveness beyond chat.
Key Capabilities
1. Computer Use
Claude can control a computer environment:
- File system operations (read, write, edit)
- Terminal command execution
- Iterative debugging and refinement
- Multi-step autonomous workflows
Use Cases: Finance agents, personal assistants, customer support, development agents, research agents
2. Built-in Tools
| Category | Tools |
|---|---|
| Files | Read, Write, Edit |
| Commands | Bash |
| Search | Grep, Glob |
| Web | WebFetch, WebSearch |
3. MCP Integration
Define custom tools via Model Context Protocol servers.
Benefits:
- Standardized tool interface
- Reusable across agents
- Enterprise data connectivity
Popular MCP Servers: GitHub, Slack, PostgreSQL, MongoDB, Stripe, Salesforce
Architecture Patterns
Pattern 1: Autonomous Task Completion
Agent completes multi-step task without intervention.
User Request → Analyze → Subtasks → Execute Tools → Iterate → Result
Pattern 2: Human-in-the-Loop
Agent proposes actions, waits for approval.
Task → Plan → Human Review → Approve? → Execute → Result
Pattern 3: Iterative Refinement
Agent retries on errors automatically.
Attempt 1 → Error → Analyze → Attempt 2 → Success
See: resources/code-examples.py for full implementations
Tool Design Best Practices
DO
- Provide tools relevant to the task
- Use clear, descriptive names
- Write detailed descriptions (Claude reads these!)
- Define strict input schemas
- Implement error handling
- Return structured outputs
DON'T
- Give agents unnecessary tools
- Use ambiguous names ("handler", "processor")
- Skip input validation
- Return raw errors without context
- Hide side effects
See: resources/code-examples.py for good/bad tool examples
MCP Integration
Connecting MCP Servers
mcp_config = {
"servers": {
"github": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"env": {"GITHUB_TOKEN": os.getenv("GITHUB_TOKEN")}
}
}
}
Full example: resources/code-examples.py
Production Best Practices
1. Streaming
Show real-time progress to build user trust.
2. Error Handling
- Catch API errors, rate limits, tool failures
- Implement fallbacks and retries
- Log errors for debugging
3. Cost Optimization
- Use Haiku for simple tasks, Sonnet for complex
- Cache repetitive contexts
- Batch similar requests
- Monitor token usage
4. Security
- Restrict file/command access
- Sanitize dangerous inputs
- Audit all agent actions
- Validate tool outputs
Implementation: resources/code-examples.py
Model Selection
| Model | Best For | Cost | Speed |
|---|---|---|---|
| claude-sonnet-4-5 | Complex reasoning, computer use, multi-step tasks | $$$ | Medium |
| claude-haiku-4 | Simple tasks, format conversions, high-throughput | $ | Fast |
Testing Agents
Unit Testing
Test individual tools in isolation.
Integration Testing
Test agent workflows with multiple tools.
Evaluation Framework
Measure accuracy, latency, tool efficiency.
Examples: resources/code-examples.py
Monitoring Metrics
| Metric | Description |
|---|---|
| Tool Call Success Rate | % of tool invocations succeeding |
| Task Completion Rate | % of requests fully resolved |
| Average Iterations | Tool calls per task |
| Latency | Time to complete requests |
| Token Usage | Input + output tokens |
| Error Rate | % of requests with errors |
Decision Framework
Use Claude SDK when:
- Building on Anthropic models
- Need computer use capabilities
- Want production-ready agent framework
- Require MCP integration
- Building autonomous agents
Consider alternatives when:
- Committed to OpenAI ecosystem → AgentKit
- Need visual agent builder → AgentKit
- Require complex state machines → LangGraph
- Want full OSS control → AutoGen/LangGraph
Resources
Documentation:
GitHub:
Key Principles
- Computer Use is Game-Changing - Leverage file/bash capabilities fully
- Tools are First-Class - Design tools as carefully as prompts
- MCP for Data - Use MCP servers for enterprise connectivity
- Stream for UX - Real-time feedback builds trust
- Security Always - Validate inputs, restrict permissions, audit
- Right Model for Task - Haiku for simple, Sonnet for complex
Build powerful, autonomous agents using Claude's cutting-edge capabilities.