huggingface/smolagents · error · AgentParsingError

No '{split_token}' token provided in your output. Your outpu

Error message

No '{split_token}' token provided in your output.
Your output:
{model_output}
. Be sure to include an action, prefaced with '{split_token}'!

What it means

CodeAgent parses model output expecting an 'Code:'/'Action:' split token before the action block. extract_action splits the model output on that token; if the token is absent (or the trailing part is empty), it raises AgentParsingError telling you to preface the action with the token.

Source

Thrown at src/smolagents/agents.py:804

        """
        return list(self._step_stream(memory_step))[-1]

    def extract_action(self, model_output: str, split_token: str) -> tuple[str, str]:
        """
        Parse action from the LLM output

        Args:
            model_output (`str`): Output of the LLM
            split_token (`str`): Separator for the action. Should match the example in the system prompt.
        """
        try:
            split = model_output.split(split_token)
            rationale, action = (
                split[-2],
                split[-1],
            )  # NOTE: using indexes starting from the end solves for when you have more than one split_token in the output
        except Exception:
            raise AgentParsingError(
                f"No '{split_token}' token provided in your output.\nYour output:\n{model_output}\n. Be sure to include an action, prefaced with '{split_token}'!",
                self.logger,
            )
        return rationale.strip(), action.strip()

    def provide_final_answer(self, task: str) -> ChatMessage:
        """
        Provide the final answer to the task, based on the logs of the agent's interactions.

        Args:
            task (`str`): Task to perform.
            images (`list[PIL.Image.Image]`, *optional*): Image(s) objects.

        Returns:
            `str`: Final answer to the task.
        """
        messages = [
            ChatMessage(

View on GitHub (pinned to 30bb116109)

Solutions

1. Retry the run; format deviations are often non-deterministic and succeed on a second attempt
2. Use a stronger instruction-following model or increase max_tokens to avoid truncation before the Code: block
3. Tighten the system/prompt template to repeat the required output format, or switch to ToolCallingAgent if the model is tuned for native tool calling
4. Catch AgentParsingError and feed the error message back to the model as a corrective next step (the error text already instructs it)

Example fix

# before (model output):
I should search for the weather then report back.
# -> AgentParsingError: No 'Code:' token provided

# after (model output):
I should search for the weather.
Code:
py
weather = get_weather("Paris")
final_answer(f"It's {weather}")

from smolagents import AgentParsingError
try:
    agent.run(task)
except AgentParsingError as e:
    # feed the parser complaint back for one corrective retry
    result = agent.run(str(e) + "\nFollow the required format exactly.")

When it happens

Trigger: A CodeAgent step where the LLM outputs only prose/thoughts without a final 'Code:' block; the model uses a variant like '```python' or 'Action:' instead of the expected token; empty or truncated output after the token causes the split/unpack to fail.

Common situations: Using a weaker or non-instruct model that ignores the ReAct format; overly chatty system prompts that bury the format instructions; token/stream truncation by max_tokens; tool-calling-formatted output in a CodeAgent.

Related errors

• Only 'markdown' is supported for a string argument to `code_
• Unsupported executor type: {self.executor_type}
• Managed agents are not yet supported with remote code execut
• Error in generating model output: {e}
• Error in code parsing: {e} Make sure to provide correct code

AI-assisted analysis of huggingface/smolagents@30bb116109 (2026-08-28). Data as JSON: /api/errors/bc196ef34f76addd. Report an issue: GitHub.