Empty thoughts with multiple questions

Version

1.4.1

Which installation method(s) does this occur on?

No response

Describe the bug.

I am trying a simple workflow, handling multiple tools, with the following tools (register.py):

from nat.builder.builder import Builder
from nat.builder.framework_enum import LLMFrameworkEnum
from nat.builder.function_info import FunctionInfo
from nat.cli.register_workflow import register_function
from nat.data_models.function import FunctionBaseConfig

class MultiplyConfig(FunctionBaseConfig, name="demo_multiply"):
    pass

@register_function(config_type=MultiplyConfig, framework_wrappers=[LLMFrameworkEnum.LANGCHAIN])
async def multiply_function(_config: MultiplyConfig, _builder: Builder):
    """Register the multiply tool."""

    async def multiply(a: float, b: float) -> float:
        """Multiply two numbers together."""
        return a * b

    yield FunctionInfo.from_fn(multiply, description="Multiply two numbers together")

class WeatherConfig(FunctionBaseConfig, name="demo_weather"):
    pass

@register_function(config_type=WeatherConfig, framework_wrappers=[LLMFrameworkEnum.LANGCHAIN])
async def weather_function(_config: WeatherConfig, _builder: Builder):
    """Register the weather tool (fake data for demo)."""

    async def weather(city: str) -> str:
        """Get the current weather for a given city."""
        fake_data = {
            "london": "12°C, cloudy",
            "new york": "8°C, sunny",
            "tokyo": "18°C, light rain",
            "paris": "14°C, partly cloudy",
        }
        return fake_data.get(city.lower(), f"22°C, clear skies in {city}")

    yield FunctionInfo.from_fn(weather, description="Get the current weather for a given city")

and the following config (workflow.yml):

llms:
  my_llm:
    _type: openai
    base_url: "https://inference-api.nvidia.com/v1"
    model_name: nvidia/nvidia/Nemotron-3-Nano-30B-A3B
    temperature: 0.0
    max_tokens: 1024

functions:
  multiply:
    _type: demo_multiply
  weather:
    _type: demo_weather

# Simple single-agent version for testing
workflow:
  _type: react_agent
  tool_names: [multiply, weather]
  llm_name: my_llm
  verbose: true
  parse_agent_response_max_retries: 3

If I ask one question, this works well:

$ nat run --config_file nat_agent/src/nat_demo_agent/configs/workflow-simple.yml --input "What is 12 times 11?"
2026-02-18 08:57:23 - INFO     - nat.cli.commands.start:192 - Starting NAT from config file: 'nat_agent/src/nat_demo_agent/configs/workflow-simple.yml'

Configuration Summary:
--------------------
Workflow Type: react_agent
Number of Functions: 2
Number of Function Groups: 0
Number of LLMs: 1
Number of Embedders: 0
Number of Memory: 0
Number of Object Stores: 0
Number of Retrievers: 0
Number of TTC Strategies: 0
Number of Authentication Providers: 0

2026-02-18 08:57:24 - INFO     - nat.runtime.session:298 - Shared workflow built (entry_function=None)
2026-02-18 08:57:25 - INFO     - httpx:1740 - HTTP Request: POST https://inference-api.nvidia.com/v1/chat/completions "HTTP/1.1 200 OK"
2026-02-18 08:57:25 - INFO     - nat.agent.react_agent.agent:171 - 
------------------------------
[AGENT]
Agent input: What is 12 times 11?
Agent's thoughts: 
r.
</think>
Question: What is 12 times 11?
Thought: I need to multiply 12 by 11 to get the answer.
Action: multiply
Action Input: {'a': 12, 'b': 11}

------------------------------
2026-02-18 08:57:25 - INFO     - nat.agent.base:221 - 
------------------------------
[AGENT]
Calling tools: multiply
Tool's input: {'a': 12, 'b': 11}
Tool's response: 
132.0
------------------------------
2026-02-18 08:57:25 - INFO     - httpx:1740 - HTTP Request: POST https://inference-api.nvidia.com/v1/chat/completions "HTTP/1.1 200 OK"
2026-02-18 08:57:25 - INFO     - nat.agent.react_agent.agent:198 - 
------------------------------
[AGENT]
Agent input: What is 12 times 11?
Agent's thoughts: 
2.
</think>
Thought: I now know the final answer
Final Answer: 132
------------------------------
2026-02-18 08:57:25 - INFO     - nat.front_ends.console.console_front_end_plugin:104 - --------------------------------------------------
Workflow Result:
['132']
--------------------------------------------------

If I ask two questions, it also works:

$ nat run --config_file nat_agent/src/nat_demo_agent/configs/workflow-simple.yml --input "What is 12 times 11? what's the weather in Tokyo?"
2026-02-18 08:58:07 - INFO     - nat.cli.commands.start:192 - Starting NAT from config file: 'nat_agent/src/nat_demo_agent/configs/workflow-simple.yml'

Configuration Summary:
--------------------
Workflow Type: react_agent
Number of Functions: 2
Number of Function Groups: 0
Number of LLMs: 1
Number of Embedders: 0
Number of Memory: 0
Number of Object Stores: 0
Number of Retrievers: 0
Number of TTC Strategies: 0
Number of Authentication Providers: 0

2026-02-18 08:58:07 - INFO     - nat.runtime.session:298 - Shared workflow built (entry_function=None)
2026-02-18 08:58:08 - INFO     - httpx:1740 - HTTP Request: POST https://inference-api.nvidia.com/v1/chat/completions "HTTP/1.1 200 OK"
2026-02-18 08:58:08 - INFO     - nat.agent.react_agent.agent:171 - 
------------------------------
[AGENT]
Agent input: What is 12 times 11? what's the weather in Tokyo?
Agent's thoughts: 
n.
</think>
Question: What is 12 times 11?  
Thought: I need to compute the product of 12 and 11 using the multiply tool.  
Action: multiply  
Action Input: {'a': 12, 'b': 11}
------------------------------
2026-02-18 08:58:08 - INFO     - nat.agent.base:221 - 
------------------------------
[AGENT]
Calling tools: multiply
Tool's input: {'a': 12, 'b': 11}
Tool's response: 
132.0
------------------------------
2026-02-18 08:58:09 - INFO     - httpx:1740 - HTTP Request: POST https://inference-api.nvidia.com/v1/chat/completions "HTTP/1.1 200 OK"
2026-02-18 08:58:09 - INFO     - nat.agent.react_agent.agent:198 - 
------------------------------
[AGENT]
Agent input: What is 12 times 11? what's the weather in Tokyo?
Agent's thoughts: 
o.
</think>
Question: what's the weather in Tokyo?  
Thought: I need to retrieve the current weather for Tokyo using the weather tool.  
Action: weather  
Action Input: {'city': 'Tokyo'}
------------------------------
2026-02-18 08:58:09 - INFO     - nat.agent.base:221 - 
------------------------------
[AGENT]
Calling tools: weather
Tool's input: {'city': 'Tokyo'}
Tool's response: 
18°C, light rain
------------------------------
2026-02-18 08:58:09 - INFO     - httpx:1740 - HTTP Request: POST https://inference-api.nvidia.com/v1/chat/completions "HTTP/1.1 200 OK"
2026-02-18 08:58:09 - INFO     - nat.agent.react_agent.agent:198 - 
------------------------------
[AGENT]
Agent input: What is 12 times 11? what's the weather in Tokyo?
Agent's thoughts: 
t.
</think>
Thought: I now know the final answer  
Final Answer: 12 times 11 equals **132.0**, and the current weather in Tokyo is **18°C with light rain**.
------------------------------
2026-02-18 08:58:09 - INFO     - nat.front_ends.console.console_front_end_plugin:104 - --------------------------------------------------
Workflow Result:
['12 times 11 equals **132.0**, and the current weather in Tokyo is **18°C with light rain**.']
--------------------------------------------------

If I ask 3 questions, it fails with an empty thought:

$ nat run --config_file nat_agent/src/nat_demo_agent/configs/workflow-simple.yml --input "What's 12*11? What's 10*10? What's the weather in Tokyo?"
2026-02-18 09:03:11 - INFO     - nat.cli.commands.start:192 - Starting NAT from config file: 'nat_agent/src/nat_demo_agent/configs/workflow-simple.yml'

Configuration Summary:
--------------------
Workflow Type: react_agent
Number of Functions: 2
Number of Function Groups: 0
Number of LLMs: 1
Number of Embedders: 0
Number of Memory: 0
Number of Object Stores: 0
Number of Retrievers: 0
Number of TTC Strategies: 0
Number of Authentication Providers: 0

2026-02-18 09:03:12 - INFO     - nat.runtime.session:298 - Shared workflow built (entry_function=None)
2026-02-18 09:03:12 - INFO     - httpx:1740 - HTTP Request: POST https://inference-api.nvidia.com/v1/chat/completions "HTTP/1.1 200 OK"
2026-02-18 09:03:13 - INFO     - nat.agent.react_agent.agent:171 - 
------------------------------
[AGENT]
Agent input: What's 12*11? What's 10*10? What's the weather in Tokyo?
Agent's thoughts: 

------------------------------
2026-02-18 09:03:13 - INFO     - nat.agent.react_agent.agent:242 - [AGENT] Retrying ReAct Agent, including output parsing Observation
2026-02-18 09:03:14 - INFO     - httpx:1740 - HTTP Request: POST https://inference-api.nvidia.com/v1/chat/completions "HTTP/1.1 400 Bad Request"
2026-02-18 09:03:14 - ERROR    - nat.agent.react_agent.agent:246 - [AGENT] Failed to call agent_node: Error code: 400 - {'error': {'message': 'litellm.BadRequestError: OpenAIException - {"error":{"message":"request: Value error, message content cannot be empty","type":"BadRequestError","code":400}}. Received Model Group=nvidia/nvidia/Nemotron-3-Nano-30B-A3B\nAvailable Model Group Fallbacks=None', 'type': None, 'param': None, 'code': '400'}}
2026-02-18 09:03:14 - ERROR    - nat.agent.react_agent.register:165 - [AGENT] ReAct Agent failed with exception: Error code: 400 - {'error': {'message': 'litellm.BadRequestError: OpenAIException - {"error":{"message":"request: Value error, message content cannot be empty","type":"BadRequestError","code":400}}. Received Model Group=nvidia/nvidia/Nemotron-3-Nano-30B-A3B\nAvailable Model Group Fallbacks=None', 'type': None, 'param': None, 'code': '400'}}
2026-02-18 09:03:14 - ERROR    - nat.builder.function:206 - Error with ainvoke in function with input: What's 12*11? What's 10*10? What's the weather in Tokyo?. Error: Error code: 400 - {'error': {'message': 'litellm.BadRequestError: OpenAIException - {"error":{"message":"request: Value error, message content cannot be empty","type":"BadRequestError","code":400}}. Received Model Group=nvidia/nvidia/Nemotron-3-Nano-30B-A3B\nAvailable Model Group Fallbacks=None', 'type': None, 'param': None, 'code': '400'}}
2026-02-18 09:03:14 - ERROR    - nat.runtime.runner:231 - Error running workflow: Error code: 400 - {'error': {'message': 'litellm.BadRequestError: OpenAIException - {"error":{"message":"request: Value error, message content cannot be empty","type":"BadRequestError","code":400}}. Received Model Group=nvidia/nvidia/Nemotron-3-Nano-30B-A3B\nAvailable Model Group Fallbacks=None', 'type': None, 'param': None, 'code': '400'}}
2026-02-18 09:03:14 - ERROR    - nat.cli.commands.start:243 - Failed to initialize workflow
Error: Error code: 400 - {'error': {'message': 'litellm.BadRequestError: OpenAIException - {"error":{"message":"request: Value error, message content cannot be empty","type":"BadRequestError","code":400}}. Received Model Group=nvidia/nvidia/Nemotron-3-Nano-30B-A3B\nAvailable Model Group Fallbacks=None', 'type': None, 'param': None, 'code': '400'}}

... but not always:

$ nat run --config_file nat_agent/src/nat_demo_agent/configs/workflow-simple.yml --input "What is 12 times 11? what's the weather in Tokyo? what is the weather in Paris?"
2026-02-18 09:04:05 - INFO     - nat.cli.commands.start:192 - Starting NAT from config file: 'nat_agent/src/nat_demo_agent/configs/workflow-simple.yml'

Configuration Summary:
--------------------
Workflow Type: react_agent
Number of Functions: 2
Number of Function Groups: 0
Number of LLMs: 1
Number of Embedders: 0
Number of Memory: 0
Number of Object Stores: 0
Number of Retrievers: 0
Number of TTC Strategies: 0
Number of Authentication Providers: 0

2026-02-18 09:04:05 - INFO     - nat.runtime.session:298 - Shared workflow built (entry_function=None)
2026-02-18 09:04:06 - INFO     - httpx:1740 - HTTP Request: POST https://inference-api.nvidia.com/v1/chat/completions "HTTP/1.1 200 OK"
2026-02-18 09:04:06 - INFO     - nat.agent.react_agent.agent:171 - 
------------------------------
[AGENT]
Agent input: What is 12 times 11? what's the weather in Tokyo? what is the weather in Paris?
Agent's thoughts: 
y.
</think>
Question: What is 12 times 11? what's the weather in Tokyo? what is the weather in Paris?
Thought: I need to compute 12 × 11 first, then get the current weather for Tokyo and Paris. I'll start with the multiplication.
Action: multiply
Action Input: {'a': 12, 'b': 11}

------------------------------
2026-02-18 09:04:06 - INFO     - nat.agent.base:221 - 
------------------------------
[AGENT]
Calling tools: multiply
Tool's input: {'a': 12, 'b': 11}
Tool's response: 
132.0
------------------------------
2026-02-18 09:04:06 - INFO     - httpx:1740 - HTTP Request: POST https://inference-api.nvidia.com/v1/chat/completions "HTTP/1.1 200 OK"
2026-02-18 09:04:06 - INFO     - nat.agent.react_agent.agent:198 - 
------------------------------
[AGENT]
Agent input: What is 12 times 11? what's the weather in Tokyo? what is the weather in Paris?
Agent's thoughts: 
t.
</think>
Question: What is the current weather in Tokyo?
Thought: I need to retrieve the current weather for Tokyo using the weather tool.
Action: weather
Action Input: {'city': 'Tokyo'}
------------------------------
2026-02-18 09:04:06 - INFO     - nat.agent.base:221 - 
------------------------------
[AGENT]
Calling tools: weather
Tool's input: {'city': 'Tokyo'}
Tool's response: 
18°C, light rain
------------------------------
2026-02-18 09:04:08 - INFO     - httpx:1740 - HTTP Request: POST https://inference-api.nvidia.com/v1/chat/completions "HTTP/1.1 200 OK"
2026-02-18 09:04:08 - INFO     - nat.agent.react_agent.agent:198 - 
------------------------------
[AGENT]
Agent input: What is 12 times 11? what's the weather in Tokyo? what is the weather in Paris?
Agent's thoughts: 
l.
</think>
Question: What is the current weather in Paris?
Thought: I need to retrieve the current weather for Paris using the weather tool.
Action: weather
Action Input: {'city': 'Paris'}
------------------------------
2026-02-18 09:04:08 - INFO     - nat.agent.base:221 - 
------------------------------
[AGENT]
Calling tools: weather
Tool's input: {'city': 'Paris'}
Tool's response: 
14°C, partly cloudy
------------------------------
2026-02-18 09:04:09 - INFO     - httpx:1740 - HTTP Request: POST https://inference-api.nvidia.com/v1/chat/completions "HTTP/1.1 200 OK"
2026-02-18 09:04:09 - INFO     - nat.agent.react_agent.agent:198 - 
------------------------------
[AGENT]
Agent input: What is 12 times 11? what's the weather in Tokyo? what is the weather in Paris?
Agent's thoughts: 
r.
</think>
Thought: I now know the final answer.  
Final Answer:  
- 12 times 11 equals **132**.  
- The current weather in **Tokyo** is **18 °C, light rain**.  
- The current weather in **Paris** is **14 °C, partly cloudy**.
------------------------------
2026-02-18 09:04:09 - INFO     - nat.front_ends.console.console_front_end_plugin:104 - --------------------------------------------------
Workflow Result:
['- 12 times 11 equals **132**.  \n- The current weather in **Tokyo** is **18\u202f°C, light rain**.  \n- The current weather in **Paris** is **14\u202f°C, partly cloudy**.']

The "funny" things is that it works with "Paris" but not "paris", never with London, whatever the case.

I have tried different things (system prompt, playing with additional parameters, etc.) with no success.

Minimum reproducible example

Relevant log output

No response

Other/Misc.

No response

Code of Conduct

• I agree to follow the NeMo Agent Toolkit Code of Conduct
• I have searched the open bugs and have found no duplicates for this bug report

Start with the react agent entry points shown in nat.agent.react_agent.agent and reproduce the three-question command using workflow-simple.yml, register.py, and the listed workflow.yml configuration. Trace the empty thought and retry path around the logged agent_node and output-parsing messages. Done means the same multi-question input no longer produces an empty message or the shown 400 error.