2nd/10_Wiki/Topics/AI_and_ML/Tool-Usage-Optimization.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-tool-usage-optimization	Tool Usage Optimization	10_Wiki/Topics	verified	self	Function Calling Tool Use Agent Tools	none	A	0.9	applied	llm tool-use function-calling agentic		2026-05-10	pending	language framework Python Anthropic SDK / OpenAI SDK / MCP

Tool Usage Optimization

매 한 줄

"매 LLM의 외부 capability 의 invoke". Tool use (function calling) 매 LLM 이 external API/code/DB 의 structured call 하는 패러다임. 2023 OpenAI function calling → 2024 parallel tools → 2025 MCP standard → 2026 매 every agent 의 backbone. 매 quality 매 tool 정의의 crisp / parallel call / error handling / cache strategy 에 의해 결정.

매 핵심

매 핵심 mechanics

• Schema — 매 JSON Schema 로 tool input 정의 — 매 model 이 fill.
• Choice — auto / tool (forced) / none.
• Parallel — 매 single turn 에 multiple tool calls (Claude / GPT-4o+).
• Iterative loop — 매 tool result 의 feed back → continue → repeat.

매 design principles

• Crisp descriptions — 매 tool description 매 "use this when X, not when Y" 명시.
• Few well-named tools > 매 many overlapping — 매 selection error rate 의 reduce.
• Idempotent / safe — 매 retry-safe — 매 LLM 매 retry 함.
• Structured errors — 매 tool error 매 LLM 이 recover 할 수 있게 actionable.

매 modern (2025-2026)

• MCP (Model Context Protocol) — 매 Anthropic 이 2024-11 release. 매 client/server tool sharing standard. 매 Claude Desktop, Cursor, Windsurf 모두 support.
• Tool result caching — 매 expensive tool (DB query, web fetch) 매 cache + hash check.
• Tool budget — 매 agent 의 max-call limit 의 prevent runaway.

매 응용

1. Code agents (file ops, shell, search).
2. Customer support (ticket, KB, account API).
3. Data analysis (SQL, plot, fetch).

💻 패턴

Anthropic tool definition

tools = [{
    "name": "get_weather",
    "description": "Get current weather. Use ONLY for weather; do not use for forecasts beyond 24h.",
    "input_schema": {
        "type": "object",
        "properties": {
            "location": {"type": "string", "description": "City, country (e.g. 'Seoul, KR')"},
            "unit": {"type": "string", "enum": ["c", "f"], "default": "c"},
        },
        "required": ["location"],
    },
}]

resp = client.messages.create(
    model="claude-opus-4-7",
    max_tokens=2000,
    tools=tools,
    messages=[{"role": "user", "content": "Weather in Seoul?"}],
)

Parallel tool execution

import asyncio

async def run_tool(call):
    return await TOOLS[call.name](**call.input)

# 매 LLM이 매 multiple tool_use blocks 의 emit
calls = [b for b in resp.content if b.type == "tool_use"]
results = await asyncio.gather(*(run_tool(c) for c in calls))
# 매 모든 result 의 single user message 로 feed back

Tool loop (agentic)

messages = [{"role": "user", "content": user_query}]
for _ in range(MAX_TURNS):
    resp = client.messages.create(model="claude-opus-4-7", tools=tools, messages=messages)
    messages.append({"role": "assistant", "content": resp.content})
    if resp.stop_reason != "tool_use":
        break
    tool_results = [
        {"type": "tool_result", "tool_use_id": b.id,
         "content": str(TOOLS[b.name](**b.input))}
        for b in resp.content if b.type == "tool_use"
    ]
    messages.append({"role": "user", "content": tool_results})

Structured error

def search_db(query: str):
    try:
        return {"ok": True, "rows": db.execute(query).fetchall()}
    except SQLError as e:
        return {"ok": False, "error": str(e),
                "hint": "Check column names; use `\\d table` to inspect."}

Result caching

import hashlib, json, functools

@functools.lru_cache(maxsize=1024)
def _cached_fetch(url_hash: str, url: str):
    return requests.get(url, timeout=10).text

def web_fetch(url: str):
    return _cached_fetch(hashlib.sha256(url.encode()).hexdigest(), url)

Prompt cache + tools (Anthropic)

resp = client.messages.create(
    model="claude-opus-4-7",
    tools=tools,                       # 매 tools 의 cacheable
    system=[{"type": "text", "text": SYSTEM, "cache_control": {"type": "ephemeral"}}],
    messages=messages,
)

MCP server (Python)

from mcp.server.fastmcp import FastMCP
mcp = FastMCP("my-tools")

@mcp.tool()
def query_orders(customer_id: str) -> list[dict]:
    """Return recent orders for a customer."""
    return db.fetch_orders(customer_id)

if __name__ == "__main__":
    mcp.run()

매 결정 기준

상황	Approach
Few tools, in-app	Direct SDK tools
Many shared tools across apps	MCP server
Latency-critical	Pre-fetch / parallel tools
Expensive tools	Cache by input hash
Untrusted LLM output	Validate + sandbox tool exec
Long agent loop	Tool budget + checkpoint

기본값: 매 5-15 well-named tools, parallel calling enabled, structured errors, prompt caching on system + tool definitions.

🔗 Graph

• 변형: Function-Calling · MCP · ReAct
• 응용: Code-Agent
• Adjacent: RAG · Structured-Output

🤖 LLM 활용

언제: 매 external state / capability 의 필요한 모든 LLM app — search, DB, email, code exec, API call. 언제 X: 매 pure text generation (summary, translate) — 매 tool 매 unnecessary.

❌ 안티패턴

• Vague tool names: "do_thing" — 매 LLM 매 selection 의 fail.
• Too many tools: 매 30+ tool 매 confuse — 매 group / route.
• No retry on transient error: 매 502 / timeout 매 transient — 매 retry-safe + LLM 의 hint.
• Streaming tool output to user mid-call: 매 tool result 매 internal — final assistant text 만 user 에.
• No max-turn limit: 매 infinite loop 의 risk.

🧪 검증 / 중복

• Verified (Anthropic tool use docs, OpenAI function calling, MCP spec 2024-11, parallel tool calling).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — modern tool use + MCP + caching patterns