Bug: OpenAI models on Bedrock don't execute tools due to stopReason handling

[jsamuel1]

Bug: OpenAI models on Bedrock don't execute tools due to stopReason handling

Summary

OpenAI models on Bedrock return stopReason: "end_turn" even when making tool calls, but Strands only executes tools when stopReason: "tool_use". This causes tool calls to be displayed but never executed.

Environment

• Strands Agents version: Latest
• Model: openai.gpt-oss-120b-1:0 on AWS Bedrock
• Working model: us.anthropic.claude-sonnet-4-20250514-v1:0

Reproduction Code

#!/usr/bin/env python3
import asyncio
import os
from strands import Agent
from strands.models import BedrockModel
from strands_tools.current_time import current_time

def create_agent():
    model = BedrockModel(
        model_id="openai.gpt-oss-120b-1:0",  # Broken
        # model_id="us.anthropic.claude-sonnet-4-20250514-v1:0",  # Works
        max_tokens=1000,
        temperature=0.2
    )
    
    return Agent(
        model=model,
        system_prompt="You are a helpful assistant.",
        tools=[current_time]
    )

async def test_tool_execution():
    agent = create_agent()
    async for event in agent.stream_async("what time is it?"):
        if "message" in event:
            print("Event:", event)

asyncio.run(test_tool_execution())

Expected Behavior

1. Model makes tool call
2. Tool is executed
3. Tool result is returned
4. Model generates final response

Actual Behavior

1. Model makes tool call ✓
2. Tool execution is skipped ❌
3. Conversation ends without tool result

Root Cause Analysis

OpenAI Response (Broken)

{
  "output": {
    "message": {
      "content": [
        {"toolUse": {"name": "current_time", "toolUseId": "tooluse_123", "input": {}}}
      ]
    }
  },
  "stopReason": "end_turn"  // ← Problem: should trigger tool execution
}

Claude Response (Working)

{
  "output": {
    "message": {
      "content": [
        {"toolUse": {"name": "current_time", "toolUseId": "tooluse_456", "input": {}}}
      ]
    }
  },
  "stopReason": "tool_use"  // ← Correctly triggers tool execution
}

Framework Logic Issue

In strands/event_loop/event_loop.py:

# Current logic - only executes tools for "tool_use" stop reason
if stop_reason == "tool_use":
    events = _handle_tool_execution(...)
    # Execute tools and continue
else:
    # End conversation - this happens for OpenAI's "end_turn"
    return

Suggested Fix

Change the logic to check for tool calls in message content regardless of stop reason:

# Check if message contains tool calls, regardless of stop reason
has_tool_calls = any(
    "toolUse" in content_item 
    for content_item in message.get("content", [])
)

if has_tool_calls:
    events = _handle_tool_execution(...)
    # Execute tools and continue
else:
    # End conversation only if no tools to execute
    return

Impact

• Affects all OpenAI models on Bedrock
• Tool calls are visible but never executed
• Breaks tool-based workflows for OpenAI models

Workaround

Use Claude models instead of OpenAI models on Bedrock.

Additional Context

This appears to be a difference in how OpenAI vs Anthropic models signal tool usage completion. The framework should handle both patterns consistently.

[jsamuel1]

Self-Contained Test Case

Here's a complete test that demonstrates the exact issue with raw Bedrock responses:

#!/usr/bin/env python3
"""
Simple test to demonstrate the stopReason difference between OpenAI and Claude models.
This test shows the exact stopReason returned by each model when making tool calls.
"""

import asyncio
import logging
from strands import Agent
from strands.models import BedrockModel
from strands_tools.current_time import current_time

# Enable debug logging to see raw responses
logging.basicConfig(level=logging.DEBUG)

async def test_stopReason(model_id: str, model_name: str):
    """Test a model and capture its stopReason behavior."""
    print(f"\n{'='*50}")
    print(f"Testing: {model_name}")
    print(f"Model: {model_id}")
    print(f"{'='*50}")
    
    model = BedrockModel(
        model_id=model_id,
        max_tokens=1000,
        temperature=0.2,
        streaming=False
    )
    
    agent = Agent(
        model=model,
        system_prompt="You are a helpful assistant.",
        tools=[current_time]
    )
    
    # Count events to see what happens
    message_count = 0
    tool_result_count = 0
    
    async for event in agent.stream_async("what time is it?"):
        if "message" in event:
            message_count += 1
            message = event["message"]
            
            # Check if this message has tool calls
            has_tools = any("toolUse" in item for item in message.get("content", []))
            
            if has_tools:
                print(f"✓ Message {message_count}: Contains tool call")
            else:
                print(f"• Message {message_count}: Regular message")
                
            # Check for tool results
            has_tool_results = any("toolResult" in item for item in message.get("content", []))
            if has_tool_results:
                tool_result_count += 1
                print(f"✓ Tool result found in message {message_count}")
    
    print(f"\nResults:")
    print(f"- Total messages: {message_count}")
    print(f"- Tool results: {tool_result_count}")
    print(f"- Tool execution: {'SUCCESS' if tool_result_count > 0 else 'FAILED'}")

async def main():
    """Compare OpenAI vs Claude stopReason behavior."""
    print("Strands Tool Execution Bug - stopReason Analysis")
    print("=" * 60)
    
    # Test OpenAI (broken)
    await test_stopReason(
        "openai.gpt-oss-120b-1:0",
        "OpenAI GPT OSS 120B"
    )
    
    # Test Claude (working)
    await test_stopReason(
        "us.anthropic.claude-sonnet-4-20250514-v1:0", 
        "Claude Sonnet 4"
    )
    
    print(f"\n{'='*60}")
    print("ANALYSIS")
    print(f"{'='*60}")
    print("Look at the debug logs above for 'stopReason' entries:")
    print("- OpenAI should show: stopReason='end_turn' (WRONG)")
    print("- Claude should show: stopReason='tool_use' (CORRECT)")
    print("\nThis difference causes Strands to skip tool execution for OpenAI models.")

if __name__ == "__main__":
    asyncio.run(main())

Test Results

Raw Bedrock Responses (from debug logs)

OpenAI Model Response:

{
  "output": {
    "message": {
      "content": [
        {"toolUse": {"name": "current_time", "toolUseId": "tooluse_..."}}
      ]
    }
  },
  "stopReason": "end_turn"  // ← PROBLEM: Should be "tool_use"
}

Claude Model Response:

{
  "output": {
    "message": {
      "content": [
        {"toolUse": {"name": "current_time", "toolUseId": "tooluse_..."}}
      ]
    }
  },
  "stopReason": "tool_use"  // ← CORRECT: Triggers tool execution
}

Test Output

Testing: OpenAI GPT OSS 120B
Model: openai.gpt-oss-120b-1:0
✓ Message 1: Contains tool call

Results:
- Total messages: 1
- Tool results: 0
- Tool execution: FAILED

Testing: Claude Sonnet 4
Model: us.anthropic.claude-sonnet-4-20250514-v1:0
✓ Message 1: Contains tool call
• Message 2: Regular message
✓ Tool result found in message 2
• Message 3: Regular message

Results:
- Total messages: 3
- Tool results: 1
- Tool execution: SUCCESS

Proof of Root Cause

This test definitively proves:

1. Both models make identical tool calls in their message content
2. OpenAI returns stopReason: "end_turn" → Strands skips tool execution (1 message total)
3. Claude returns stopReason: "tool_use" → Strands executes tools correctly (3 messages: call → result → response)
4. The bug is in Strands' event loop logic that only executes tools when stop_reason == "tool_use"

Environment Tested

• Strands Agents: 1.4.0 (latest)
• Both models tested with streaming=False to avoid streaming compatibility issues

[pgrayy]

Have you brought this issue up with Bedrock yet? The contract from Bedrock is that a model requesting tools needs to set stop reason to "tool_use" (docs). Strands adheres to this contract and thus I would not consider this behavior a bug in Strands.

[jsamuel1]

Bedrock team notified.
I agree - Bedrock bug, not a Strands Bug.

I suggest leaving this open as a placeholder until the bedrock bug is fixed.

[mimocha]

I had one case of this happen a few hours ago, but I was using amazon.nova-pro-v1:0.
It only happened once in a few hundred executions, and I haven't been able to reproduce the issue again so far.

Environment

• python 3.13.2
• strands-agents 1.4.0
• strands-agents-tools 0.2.3
• boto3 1.26.0

[benjaminfiliatrault]

@jsamuel1 do you have the issue link so we can track the Bedrock issue?

[LucaButBoring]

The ticket was internal (I'm not the one who handled it) - is this still reproducible? It seems to be fixed for me, now.

[pgrayy]

As an added layer of protection, we added some logic in the BedrockModelProvider to handle invalid stop reasons. Please let us know if you continue to encounter any issues.

[fvisticot]

I'm using Strands 1.7.1 with tool calling with "openai.gpt-oss-120b-1:0" and it works fine.