Xây dựng agent nghiên cứu với Web Search + URL Extract (tool use của Claude, từ đầu đến cuối)

import requests

API_KEY = "pk_yourkey"

def fetch_tool(tool_path: str) -> dict:
    """Fetch a Claude tool definition from API Pick's tool-schema endpoint."""
    schema = requests.get(f"https://www.apipick.com{tool_path}/tool-schema").json()
    return schema["claude"]

WEB_SEARCH_TOOL = fetch_tool("/api/search/web")
URL_EXTRACT_TOOL = fetch_tool("/api/extract")

# Cache these once at module load. They don't change between requests.

def call_tool(tool_use_block) -> dict:
    """Execute the tool Claude asked for and return a tool_result block."""
    name = tool_use_block.name
    args = tool_use_block.input

    if name == "web_search":
        endpoint = "https://www.apipick.com/api/search/web"
    elif name == "extract_urls":
        endpoint = "https://www.apipick.com/api/extract"
    else:
        return {
            "type": "tool_result",
            "tool_use_id": tool_use_block.id,
            "content": f"Unknown tool: {name}",
            "is_error": True,
        }

    resp = requests.post(
        endpoint,
        headers={"x-api-key": API_KEY},
        json=args,
        timeout=30,
    )

    if resp.status_code != 200:
        return {
            "type": "tool_result",
            "tool_use_id": tool_use_block.id,
            "content": f"HTTP {resp.status_code}: {resp.text[:500]}",
            "is_error": True,
        }

    return {
        "type": "tool_result",
        "tool_use_id": tool_use_block.id,
        "content": resp.text,
    }

import anthropic

client = anthropic.Anthropic()

SYSTEM_PROMPT = """You are a research assistant. When the user asks a question:

1. Use web_search to find relevant sources. Prefer recent (set start_date) for time-sensitive queries.
2. Use extract_urls to read the body of the top 3-5 most relevant URLs from the search results.
3. Answer the question concisely. After every factual claim, include an inline citation in the form [source: URL].
4. If you don't have enough information, say so — don't fabricate.

Be concise. Aim for 3-5 sentence answers unless the user asks for depth."""

def research(question: str) -> str:
    messages = [{"role": "user", "content": question}]

    while True:
        response = client.messages.create(
            model="claude-sonnet-4-6",
            max_tokens=2048,
            system=SYSTEM_PROMPT,
            tools=[WEB_SEARCH_TOOL, URL_EXTRACT_TOOL],
            messages=messages,
        )

        # Append the assistant's response (may contain tool_use blocks)
        messages.append({"role": "assistant", "content": response.content})

        if response.stop_reason == "end_turn":
            # Pull out the text answer
            return "\n".join(b.text for b in response.content if b.type == "text")

        if response.stop_reason == "tool_use":
            # Run every tool the model asked for in this turn
            tool_results = [
                call_tool(b)
                for b in response.content if b.type == "tool_use"
            ]
            messages.append({"role": "user", "content": tool_results})
            continue

        raise RuntimeError(f"Unexpected stop_reason: {response.stop_reason}")

if __name__ == "__main__":
    answer = research("What were the major announcements at OpenAI DevDay this year?")
    print(answer)

OpenAI's DevDay focused on three themes: cheaper inference (a new gpt-mini
tier at roughly half the prior cost) [source: https://openai.com/devday-2026],
agent infrastructure (a managed agent runtime with persistent memory and
tool sandboxing) [source: https://techcrunch.com/...], and developer
tooling (a new Responses API replacing the legacy Assistants API)
[source: https://platform.openai.com/docs/...].