Description

When using CrewAI with an AWS Bedrock LLM (e.g. Amazon Nova Pro via bedrock/us.amazon.nova-pro-v1:0), tool arguments are silently discarded. The LLM correctly calls the tool with the right arguments, but the tool receives an empty dict {},
causing either a TypeError (missing required positional argument) or a Pydantic ValidationError (field required).

Affected versions: Confirmed in 1.9.3 and 1.13.0.

Steps to Reproduce

1. Define a BaseTool with a required parameter:
   from crewai.tools import BaseTool
   from pydantic import BaseModel, Field
   from typing import Type

class CityInput(BaseModel):
city: str = Field(..., description="City name")

class TravelTool(BaseTool):
name: str = "get_travel_details"
description: str = "Get travel details for a city."
args_schema: Type[BaseModel] = CityInput

  def _run(self, city: str) -> str:                                                                                                                                                                                                             
      return f"Details for {city}" 

2. Assign this tool to a CrewAI agent backed by a Bedrock LLM:
   llm = LLM(model="bedrock/us.amazon.nova-pro-v1:0", temperature=0)
   agent = Agent(tools=[TravelTool()], llm=llm, ...)

3. Run the crew. The Bedrock LLM correctly calls get_travel_details with {"city": "Paris"}, but the tool receives {}.

Error in 1.9.3:
TravelDetailsTool._run() missing 1 required positional argument: 'city'

Error in 1.13.0:
Tool 'get_travel_details' arguments validation failed: 1 validation error for CityInput
city
Field required [type=missing, input_value={}, input_type=dict]

Root Cause

File: crewai/agents/crew_agent_executor.py

In _handle_native_tool_calls / _extract_tool_call_info (line varies by version), the code that extracts tool arguments handles multiple LLM response formats:

func_info = tool_call.get("function", {})
func_args = func_info.get("arguments", "{}") or tool_call.get("input", {})

Bedrock's Converse API returns tool calls in this format:
{"name": "get_travel_details", "toolUseId": "abc123", "input": {"city": "Paris"}}

There is no "function" wrapper, so func_info = {}.

func_info.get("arguments", "{}") returns the default string "{}", which is truthy.

The or short-circuits — tool_call.get("input", {"city": "Paris"}) is never evaluated.

func_args becomes "{}", and after json.loads("{}"), args_dict = {}. The actual Bedrock arguments are completely lost.

Expected behavior

The tool should be invoked with the arguments from tool_call["input"] when processing a Bedrock-format tool call response.

Screenshots/Code snippets

Change the default from "{}" to None:

Before

func_args = func_info.get("arguments", "{}") or tool_call.get("input", {})

After

func_args = func_info.get("arguments") or tool_call.get("input", {})

With None as default, func_info.get("arguments") returns None (falsy) when no "function" wrapper is present, so the or correctly falls through to tool_call.get("input", {...}).

Operating System

Ubuntu 20.04

Python Version

3.12

crewAI Version

1.13.0

crewAI Tools Version

1.13.0

Virtual Environment

Venv

Evidence

Iteration: 1
│
│ Attempt: 0
│
│ Error: Tool 'get_travel_details' arguments validation failed: 1 validation error for TravelDetailsInput
│
│ city
│
│ Field required [type=missing, input_value={}, input_type=dict]
│
│ For further information visit https://errors.pydantic.dev/2.11/v/missing
│
│ Expected arguments: {"city": {"description": "Name of the city to retrieve travel details for.", "title": "City", "type": "string"}}
│
│ Required: ["city"]

Possible Solution

Change the default from "{}" to None:

Before

func_args = func_info.get("arguments", "{}") or tool_call.get("input", {})

After

func_args = func_info.get("arguments") or tool_call.get("input", {})

With None as default, func_info.get("arguments") returns None (falsy) when no "function" wrapper is present, so the or correctly falls through to tool_call.get("input", {...}).

Additional context

NA