claude-advanced-tool-use（Claude 高级 tool USE）

Claude Advanced Tool Use

Overview

Advanced tool use provides three complementary features that address distinct bottlenecks in production systems. These features can be used independently or combined for compounding efficiency gains of 37-85%.

The Three Features

1. Tool Search (85-95% context savings)

• Dynamically discover and load tools on-demand
• Scale to 10,000+ tools without context bloat
• Two variants: Regex pattern matching or BM25 natural language queries
• Deferred loading keeps tool definitions out of context until needed

2. Programmatic Tool Calling (37% token reduction)

• Claude writes Python code that calls your tools within a sandboxed container
• Intermediate results stay out of context window
• Eliminates 19+ unnecessary inference passes on complex workflows
• Ideal for data aggregation, filtering, and multi-step orchestration

3. Tool Use Examples (72% → 90% accuracy)

• Concrete usage examples clarify ambiguous JSON schemas
• Improve parameter handling accuracy
• Express API conventions schemas cannot capture
• Reduce hallucination on optional parameters

Key Insight: Start with your biggest bottleneck, then add complementary features as needed. Not all-or-nothing.

Feature Comparison Matrix

When to Use

Tool Search:

• You have 10+ tools in your system
• Tool definitions exceed 10K tokens combined
• Tool selection accuracy degrades with large sets
• Using MCP servers (200+ tools across servers)
• Tool library grows over time

Programmatic Calling:

• Workflows with 3+ dependent tool calls
• Processing large datasets where only summaries needed
• Batch operations across multiple items
• Conditional logic based on intermediate results
• Aggregation tasks combining multiple data sources

Tool Use Examples:

• JSON schemas don't capture when to use optional parameters
• API conventions need clarification
• Parameter handling accuracy below 80%
• Reducing hallucination on complex tool inputs

Production Patterns:

• Building scalable agentic systems
• Implementing security-first tool orchestration
• Optimizing token costs in production
• Coordinating multiple tools efficiently

Quick Start: Tool Search with Deferred Loading

Python Example:

import anthropic

client = anthropic.Anthropic()

tools = [
    # Always-loaded: tool search itself
    {
        "type": "tool_search_tool_regex_20251119",
        "name": "tool_search_tool_regex"
    },
    # Always-loaded: frequently used tools (3-5 total)
    {
        "name": "get_user_info",
        "description": "Get user information by ID",
        "input_schema": {
            "type": "object",
            "properties": {
                "user_id": {"type": "string"}
            }
        }
    },
    # Deferred: specialized tools loaded on-demand
    {
        "name": "get_weather",
        "description": "Get current weather for a location",
        "input_schema": {
            "type": "object",
            "properties": {
                "location": {"type": "string"}
            }
        },
        "defer_loading": True
    },
    {
        "name": "get_forecast",
        "description": "Get weather forecast for a location and days ahead",
        "input_schema": {
            "type": "object",
            "properties": {
                "location": {"type": "string"},
                "days": {"type": "integer"}
            }
        },
        "defer_loading": True
    }
]

response = client.beta.messages.create(
    model="claude-sonnet-4-5-20250929",
    betas=["advanced-tool-use-2025-11-20"],
    max_tokens=2048,
    messages=[{
        "role": "user",
        "content": "What's the weather in San Francisco?"
    }],
    tools=tools
)

print(response.content[0].text)

What Happens:

1. Claude analyzes request and determines weather tools are needed
2. Uses tool_search_tool_regex with pattern "weather"
3. API returns tool_reference blocks pointing to matching tools
4. Tool definitions automatically expanded
5. Claude invokes get_weather with appropriate parameters
6. Only discovered tools loaded into context (85% savings)

The Compound Effect

When combined strategically, these features multiply efficiency gains:

Example: Large MCP System

• 200+ tools via MCP servers
• Tool Search: Reduces context by 85% (10K tokens → 1.5K tokens)
• Programmatic Calling: Reduces workflow tokens by 37%
• Tool Examples: Improves accuracy from 72% to 90%
• Combined Impact: ~90% token reduction + 18% accuracy improvement

Implementation Strategy:

1. Start with tool search if definitions exceed 10K tokens
2. Add programmatic calling for multi-step workflows
3. Include examples for tools with complex parameters
4. Measure efficiency gains at each step
5. Iterate based on bottleneck analysis

Production Architecture Patterns

From production codebase analysis (Source):

1. Planner + Executor

• Separate planning conversation from execution
• Planner decides what to do, executor does it with focused context
• Reduces re-planning overhead

2. Preview Then Fetch

• Return IDs/summaries first
• Fetch full documents on-demand
• Keeps context lean

3. Guard + Act

• Validate parameters before execution
• Server-side validation with actionable errors
• Security-first approach

4. Summarize Outputs

• Compress tool results before continuing
• Works well with programmatic calling
• Only summary enters context window

See references/production-patterns.md for complete implementation templates.

Performance Optimization

Identifying Your Bottleneck:

• Tool definitions > 10K tokens → Enable tool search
• Large intermediate datasets → Use programmatic calling
• Parameter confusion → Provide tool examples
• Multiple tool calls → Consider programmatic orchestration

Optimization Checklist:

• Tool descriptions under 200 characters
• Semantic keywords in descriptions
• 3-5 most-used tools always loaded
• Clear tool naming (e.g., search_customer_orders not query_db)
• Tool index cached in-session
• Ranked results returned
• Server-side validation enabled
• Tool results memoized where appropriate
• Security best practices implemented
• Observability logging with trace IDs

Integration with Context Management

Tool search and programmatic calling work seamlessly with context editing:

Tool Result Clearing (from claude-context-management):

• Server-side strategy removes older tool results chronologically
• Preserves recent N tool uses
• Can exclude specific tools (e.g., web_search)
• Works with deferred loading

Programmatic Calling (keeps results out of context):

• Tool results from code execution don't enter context window
• Only final outputs returned to Claude
• Natural synergy: fewer results + more efficient orchestration

Combined Example:

response = client.beta.messages.create(
    model="claude-opus-4-5-20251101",
    betas=["advanced-tool-use-2025-11-20", "context-management-2025-06-27"],
    tools=tools_with_deferred_loading,
    context_management={
        "edits": [{
            "type": "clear_tool_uses_20250919",
            "trigger": {"type": "input_tokens", "value": 100000},
            "keep": {"type": "tool_uses", "value": 3}
        }]
    },
    messages=messages
)

Related Skills

• anthropic-expert: Basic tool use fundamentals, MCP integration, code execution tool
• claude-context-management: Server-side tool result clearing, token optimization
• claude-cost-optimization: Efficiency tracking, ROI measurement for tool optimizations
• claude-opus-4-5-guide: Model capabilities, effort parameter impact on tool use

References

For detailed implementation patterns, see:

• references/tool-search-patterns.md: Complete tool search guide (regex + BM25 variants, deferred loading, MCP integration, 10K+ tool scalability)
• references/programmatic-tool-calling.md: Sandboxed execution patterns, allowed_callers parameter, token efficiency mechanisms, ideal use cases
• references/production-patterns.md: Architecture patterns from production codebases, security best practices, error handling, observability
• references/performance-optimization.md: Optimization strategies, caching patterns, latency reduction, efficiency metrics