langgraph/how-tos/async/

[giscus[bot]]

langgraph/how-tos/async/

Build language agents as graphs

https://langchain-ai.github.io/langgraph/how-tos/async/

[Tamara-Codes]

Thanks for this nice tutorial, it's so clean and simple. Works like magic!

[SunGaofeng]

Can you give more examples on how to define async tools? Here you showed how to define an async function to call the llm model, but the tool function is still sync. If I want to call the tools asynchronously, how to define the tool function or tool node?

Can I just define an async tool funciton like the following:

@tool
async def search(query: str):
    """Call to surf the web."""
    # This is a placeholder, but don't tell the LLM that...
     # suppose search_tool is a tool that implements ainvoke/invoke method
    results = await search_tool.ainvoke(query)
    return results

Then can I just use the ToolNode class, or should I define a custom toolnode to handle the async calling?

[joaojhgs]

@hinthornw I'm trying to use an async tool using the proper async def and @tool decorator, but whenever it's called by the tool node it seems to call it sync which causes an error, I've updated to the latest version I could (0.4.6 and prebuilt 0..2.1) but it's still happening, how can I use async tools with the tool node?

Tool definition:

Using pre-built tool node:

Langsmith tool call error:

[hinthornw]

Use ainvoke or astream rather than the synchronous call then.

It only attempts to use sync if you try to call it in a synchronous fashion.

[joaojhgs]

Use ainvoke or astream rather than the synchronous call then.

It only attempts to use sync if you try to call it in a synchronous fashion.

I'm not directly calling the tools though, the tool node does that right? Or should I call the entire graph in async manner in order for the ToolNode to call the tools async? would that break the tools that are sync instead?

[nowjean]

Use ainvoke or astream rather than the synchronous call then.

It only attempts to use sync if you try to call it in a synchronous fashion.

I'm not directly calling the tools though, the tool node does that right? Or should I call the entire graph in async manner in order for the ToolNode to call the tools async? would that break the tools that are sync instead?

@joaojhgs If you use ainvoke or astream with graph_builder, ToolNode will automatically run asynchronously. It supports both invoke and ainvoke. You can debug this!"

[joaojhgs]

Use ainvoke or astream rather than the synchronous call then.
It only attempts to use sync if you try to call it in a synchronous fashion.

I'm not directly calling the tools though, the tool node does that right? Or should I call the entire graph in async manner in order for the ToolNode to call the tools async? would that break the tools that are sync instead?

@joaojhgs If you use ainvoke or astream with graph_builder, ToolNode will automatically run asynchronously. It supports both invoke and ainvoke. You can debug this!"

Got it! I managed to get it working!

I guess it was a bit of lack of knowledge about the event loop on my side, but I assumed running the graph syncronously wouldn't necessarily mean all tools have to be run syncronously.
Thanks everyone!

[shailesh-numaai]

I am not sure i I understood the answer correctly on how to make tool async

here is my code

from playwright.async_api import async_playwright
from langchain_openai import ChatOpenAI
from langgraph.prebuilt import create_react_agent, ToolNode
from langchain_core.tools import tool
from agents.browser_agent.prompt import BROWSER_AGENT_PROMPT

# Simple global browser instance
playwright = None
browser = None
page = None

async def get_browser(): //NOT A TOOL
    """Get or create browser instance"""
    global playwright, browser, page
    
    # Check if we already have a working page
    if page is not None:
        try:
            # Test if the page is still responsive
            await page.title()
            return page
        except Exception as e:
            # Page is broken, we'll create a new one
            pass
    
    # Create new browser session
    
    # Clean up any existing resources
    if page is not None:
        try:
            await page.close()
        except:
            pass
    if browser is not None:
        try:
            await browser.close()
        except:
            pass
    if playwright is not None:
        try:
            await playwright.stop()
        except:
            pass
    
    # Create new session
    playwright = await async_playwright().start()
    browser = await playwright.chromium.launch(headless=False)
    page = await browser.new_page()
    
    return page

@tool
async def navigate_to_url(url: str) -> str:
    """Navigate to a URL"""
    try:
        page = await get_browser()
        await page.goto(url)
        title = await page.title()
        return f"Successfully navigated to {url}"
    except Exception as e:
        return f"Error navigating to {url}: {str(e)}"

@tool
async def click_element_by_text(text: str) -> str:
    """Click an element with the given text"""
    try:
        page = await get_browser()
        current_title = await page.title()
        
        element = page.locator(f"text={text}")
        count = await element.count()
        if count > 0:
            await element.first.click()
            return f"Successfully clicked element with text: {text}"
        else:
            return f"No elements found with text: {text}"
    except Exception as e:
        return f"Error clicking element with text '{text}': {str(e)}"

@tool
async def get_page_content() -> str:
    """Get current page content to see what's available"""
    try:
        page = await get_browser()
        title = await page.title()
        all_text = await page.locator("body").inner_text()
        
        return f"Page Title: {title}\nPage Text Preview: {all_text[:500]}..."
    except Exception as e:
        return f"Error getting page content: {str(e)}"

# Create ToolNode for all your async tools
browser_tool_node = ToolNode([navigate_to_url, click_element_by_text, get_page_content])

browser_agent = create_react_agent(
    model=ChatOpenAI(model="gpt-4o", temperature=0.1),
    tools=browser_tool_node,  # Use ToolNode here
    name="browser_agent",
    prompt=BROWSER_AGENT_PROMPT,
).with_config(recursion_limit=20)

When this agent is invoked via frontend stremable chat interface.... no tools are getting called/nothing happens i can see tool call in langsmith but it stucks in process no further flow. any suggestions?