dragonkid/llm-agent-tool-calling-optimization

LLM Agent Tool Calling Performance Optimization

Overview

LLM agents tend to perform iterative sequential searches when calling server-side tools, causing linear time growth. Force parallel execution through prompt constraints and explicit tool call budgeting.

When to Use

• Agent uses server-side search tools (web_search, x_search, etc.)
• Execution time is noticeably long (30-40s per round)
• Response analysis shows multi-round sequential tool calls
• Need to limit total tool calls across multiple search tools

Core Pattern

Force parallel execution through prompt constraints with explicit numeric limits:

# Add to system prompt
## Search Constraints (CRITICAL)
- Maximum N total searches: up to X tool_a + up to Y tool_b
- Plan ALL search queries upfront, then execute them in ONE parallel batch
- Do NOT add follow-up searches based on initial results
- Use OR operators to combine multiple intents into single queries

Query Consolidation

Before (4 separate searches):

• "TOKEN launch"
• "TOKEN migration"
• "TOKEN listing"
• "TOKEN partnership"

After (1 consolidated search):

• "TOKEN (launch OR migration OR listing OR partnership)"

Time Constraints

Priority: Check SDK/API parameters first.

1. SDK supports time range (e.g., from_date, to_date): Use SDK parameters directly (hard constraint).

2. SDK does not support time range: Add soft constraints in prompt:

## Time Constraints
- Focus on events from {search_from_date} onwards
- Skip/ignore any search results dated before {search_from_date}

Diagnosis

Analyze API response for these indicators:

• server_side_tools_used: Number of tool calls
• Multiple outputs blocks: Indicates multi-round calls
• reasoning_tokens: Usually not the main bottleneck

Common Mistakes

• Allowing iterative follow-up searches based on initial results
• Not setting explicit numeric limits on total tool calls
• Forgetting to use OR operators to consolidate related intents