[BUG] Gemini agents are looping multiple times, even after getting tool response

[vizsatiz]

Describe the bug
A clear and concise description of what the bug is with stack trace

2025-11-26 18:12:37.910 IST
"reply_message": "Razorpay provides a payment solution in India that allows businesses to manage both online and in-store payments. Its Point-of-Sale (POS) system enables the acceptance of offline payments via UPI, credit, and debit cards at various touchpoints, including retail counters, doorstep deliveries, and self-serve kiosks."
2025-11-26 18:12:37.910 IST
}
2025-11-26 18:12:37.910 IST

2025-11-26 18:12:37.915 IST
2025-11-26 12:42:37 | INFO | google_genai.models | be-xM9keoNJ | models.py:5012 | AFC is enabled with max remote calls: 10.
2025-11-26 18:12:39.763 IST
2025-11-26 12:42:39 | INFO | httpx | be-xM9keoNJ | _client.py:1025 | HTTP Request: POST https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent "HTTP/1.1 200 OK"
2025-11-26 18:12:39.767 IST
2025-11-26 12:42:39 | WARNING | google_genai.types | be-xM9keoNJ | types.py:6029 | Warning: there are non-text parts in the response: ['function_call', 'thought_signature'], returning concatenated text result from text parts. Check the full candidates.content.parts accessor to get the full model response.
2025-11-26 18:12:39.767 IST
2025-11-26 12:42:39 | INFO | flo_ai | partial_tool.py:52 | Executing partial tool querying_knowlegebase with merged params: {'kb_id': 'ee3b887d-3f6e-48d4-9752-ef1bc0ee5827', 'inference_id': '55cbe356-6142-4c9f-95d8-b924e7a35ee9', 'question': 'what is razorpay'}
2025-11-26 18:12:41.637 IST
2025-11-26 12:42:41 | INFO | google_genai.models | be-xM9keoNJ | models.py:5012 | AFC is enabled with max remote calls: 10.
2025-11-26 18:12:50.858 IST
2025-11-26 12:42:50 | INFO | httpx | be-xM9keoNJ | _client.py:1025 | HTTP Request: POST https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-pro:generateContent "HTTP/1.1 200 OK"
2025-11-26 18:12:50.861 IST
2025-11-26 12:42:50 | INFO | flo_ai | partial_tool.py:56 | Partial tool querying_knowlegebase returned: ```json
2025-11-26 18:12:50.861 IST
{
2025-11-26 18:12:50.861 IST
"reply_message": "Based on the documents, Razorpay POS is a point-of-sale payment solution available in India that allows businesses to accept offline payments using UPI, credit, or debit cards. It can be used for in-person purchases, customer doorstep deliveries, unattended kiosks, and vending machines."
2025-11-26 18:12:50.861 IST
}
2025-11-26 18:12:50.861 IST

2025-11-26 18:12:50.865 IST
2025-11-26 12:42:50 | INFO | google_genai.models | be-xM9keoNJ | models.py:5012 | AFC is enabled with max remote calls: 10.
2025-11-26 18:12:53.111 IST
2025-11-26 12:42:53 | INFO | httpx | be-xM9keoNJ | _client.py:1025 | HTTP Request: POST https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash:generateContent "HTTP/1.1 200 OK"
2025-11-26 18:12:53.113 IST
2025-11-26 12:42:53 | INFO | staging-floware | be-xM9keoNJ | agent_inference_service.py:151 | Successfully completed inference for agent test-translator in 70.67 seconds
2025-11-26 18:12:53.114 IST
2025-11-26 12:42:53 | INFO | staging-floware | be-xM9keoNJ | agent_controller.py:214 | Successfully completed v2 inference for agent_id: ad0bc608-fe1f-4193-b957-19fb52cb1c4b

[harikapadia999]

Analysis: Gemini Agent Looping Issue

I've reviewed this bug and it appears to be a classic agent loop problem where the LLM continues making tool calls even after receiving responses. This is a common issue with Gemini models when tool calling isn't properly terminated.

Potential Root Causes:

1. Missing Stop Condition
The agent might not be detecting when the task is complete. Check if the agent has proper termination logic:

if tool_response_received and task_complete:
    break  # Exit the loop

2. Tool Response Format Mismatch
Gemini might not be recognizing the tool response format. Ensure responses match Gemini's expected schema:

{
    "role": "function",
    "name": "tool_name",
    "content": "tool_result"
}

3. Max Iterations Not Enforced
Related to issue #141, the max_tool_count might not be properly enforced for Gemini agents.

Debugging Steps:

1. Add Logging: Log each iteration to see what's triggering the loop

logger.debug(f"Iteration {iteration_count}: Tool={tool_name}, Response={response}")

2. Check Tool Call History: Verify if the same tool is being called repeatedly

if tool_name in recent_tool_calls:
    logger.warning(f"Tool {tool_name} called multiple times")

3. Inspect Gemini Response: Check if Gemini is returning a finish_reason

if response.candidates[0].finish_reason == "STOP":
    break

Suggested Fix:

class GeminiAgent:
    def run(self, max_iterations=10):
        iteration = 0
        while iteration < max_iterations:
            response = self.gemini_client.generate_content(...)
            
            # Check for stop condition
            if response.candidates[0].finish_reason == "STOP":
                break
                
            # Process tool calls
            if response.candidates[0].content.parts:
                for part in response.candidates[0].content.parts:
                    if hasattr(part, 'function_call'):
                        # Execute tool
                        result = self.execute_tool(part.function_call)
                        # Add result to conversation
                        self.add_tool_response(result)
            
            iteration += 1
            
        if iteration >= max_iterations:
            logger.warning("Max iterations reached, stopping agent")

Questions:

1. What's the exact tool being called repeatedly?
2. Is this happening with all Gemini models or specific ones (e.g., gemini-pro vs gemini-1.5-pro)?
3. Does the issue occur with simple tools or only complex multi-step workflows?

Would be happy to help debug further if you can share more details about the specific scenario!

[vizsatiz]

@harikapadia999 Thank you for the breakdown, we will work on this and fix it in the on going sprint