Why is my agent sometimes call tool with wrong parameter?

CrewAI Community Support
1

I use several tools to get the stock information.
But many time my agent fail to call my tool with correct paramters.

@tool("Stock Info")
def stock_info(ticker: str):
    """
    Useful to get stock infomation.(e.g. market cap, beta, 52-week high, 52-week low, etc.)
    
    Input paramter:
    - ticker(type:str): The ticker of a company.
    """
    ticker = yf.Ticker(ticker)
    return ticker.info

The tool is very simple tool with only one string paramter.
But my agent trying to use the tool with wrong parameter again and again.

Tool Used: Stock Info
  
Tool input: [{"ticker": "VST"}, {"tool_code": "Stock Info", "tool_input": {"ticker": "VST"}}]
  
Message: Thought:Okay, let's start by getting the stock information for Vistra Corp (VST) using the 'Stock Info' tool to get some initial parameters. Action: Stock Info Action Input: {"ticker": "VST"}
  
tool_code
{"tool_code": "Stock Info", "tool_input": {"ticker": "VST"}}
  
Observation: Error: the Action Input is not a valid key, value dictionary.
  
Error: the Action Input is not a valid key, value dictionary.

I’m not sure why the agent use tool input with
[{“ticker”: “VST”}, {“tool_code”: “Stock Info”, “tool_input”: {“ticker”: “VST”}}]
correct parameter is just {“ticker”:“VST”}
This happening often for me.
How can I make my agent always use correct parameters to a specified tool?
Thanks.

2

@iam1492 I recommend you set up a custom tool by subclassing the BaseTool class instead of using the @tool decorator as follows:

from typing import Type
from crewai.tools import BaseTool
from pydantic import BaseModel, Field

class MyToolInput(BaseModel):
    """Input schema for MyCustomTool."""
    argument: str = Field(..., description="Description of the argument.")

class MyCustomTool(BaseTool):
    name: str = "Name of my tool"
    description: str = "What this tool does. It's vital for effective utilization."
    args_schema: Type[BaseModel] = MyToolInput

    def _run(self, argument: str) -> str:
        # Your tool's logic here
        return "Tool's result"
1 Like
3

Hi rokbenko.
I modify one of my tool with subclassing the BaseTool class.
If this work well, I will modify all the tool in same way!
Thanks!

4

@iam1492 Please let me know if this solves the issue!

5

I found another solution with seperate llm and function_calling_llm with different model.
I use gemini 2.0 for llm but change to use ollama/hengwen/watt-tool-8B:latest for my function_calling_llm.
You can find out function calling benchmark here: Berkeley Function Calling Leaderboard V3 (aka Berkeley Tool Calling Leaderboard V3)

ollama/hengwen/watt-tool-8B is quite small llm so that I can run this model locally but have strong function.

And also this model works great without failing any function call.
Thanks

6

After doing some effort but still no help with function calling.

  • Using different llm for function calling(tool) help better tool usage, but still have same issue.
  • Change @tool decorator with subclassing the BaseTool is not working.
    Strange thing is one of my agent make this error much more than other agents but I don’t know what is problem. I will comment here when I find any solution.
7

Here I report my bug :slight_smile:

8

Can you share your full code in repo so we can clone and help troubleshoot this for you?

Also, I’ve personally noticed that most small OSS models struggle quite a bit with tool calling even if they’re high on the leaderboard. Sometimes they get it right sometimes they don’t; non-deterministic.

9

Hello :slight_smile:
My repo is currently private repository.
I use gemini-2.0 pro exp model for testing my project.
And I understanding llms sometimes confusing about using function call(tool).
What I found was:
More error occurs when I use more tool for one agent.
When I reducing tools which is provided to my agent, the agent do the mistake less.
I think this error can be disappears when model become stable later and Devins(AI coder) workaround may help until then.

10

I too observed the same, especially when we have scenario which involves more than 2 tools I’m getting this issue.

More Over I couldn’t able to understand on @Tool Decorators, and I fall back to use Subclassing the Base Tool, it didn’t help much. will look out this thread to find the solution.