huggingface/smolagents · error · AgentParsingError

Error in code parsing: {e} Make sure to provide correct code

Error message

Error in code parsing:
{e}
Make sure to provide correct code blobs.

What it means

After generation, CodeAgent._step_stream extracts executable Python from the model output — via json.loads for structured outputs or parse_code_blobs for tagged code blocks — and wraps any failure in AgentParsingError with guidance to provide correct code blobs. It means the model's textual output was not in the expected code format.

Source

Thrown at src/smolagents/agents.py:1713

                    memory_step.model_output_message.content = output_text

            memory_step.token_usage = chat_message.token_usage
            memory_step.model_output = output_text
        except Exception as e:
            raise AgentGenerationError(f"Error in generating model output:\n{e}", self.logger) from e

        ### Parse output ###
        try:
            if self._use_structured_outputs_internally:
                code_action = json.loads(output_text)["code"]
                code_action = extract_code_from_text(code_action, self.code_block_tags) or code_action
            else:
                code_action = parse_code_blobs(output_text, self.code_block_tags)
            code_action = fix_final_answer_code(code_action)
            memory_step.code_action = code_action
        except Exception as e:
            error_msg = f"Error in code parsing:\n{e}\nMake sure to provide correct code blobs."
            raise AgentParsingError(error_msg, self.logger)

        tool_call = ToolCall(
            name="python_interpreter",
            arguments=code_action,
            id=f"call_{len(self.memory.steps)}",
        )
        yield tool_call
        memory_step.tool_calls = [tool_call]

        ### Execute action ###
        self.logger.log_code(title="Executing parsed code:", content=code_action, level=LogLevel.INFO)
        try:
            code_output = self.python_executor(code_action)
            execution_outputs_console = []
            if len(code_output.logs) > 0:
                execution_outputs_console += [
                    Text("Execution logs:", style="bold"),
                    Text(code_output.logs),

View on GitHub (pinned to 30bb116109)

Solutions

Ensure your prompt templates instruct code in the same tags configured via code_block_tags (use 'markdown' with the default code-agent.yaml templates).
Retry the run — the parsing error is fed back and models usually correct the format.
Use a stronger model or enable structured outputs so the code field is guaranteed.

Example fix

# before
agent = CodeAgent(model=weak_model, tools=[], code_block_tags='markdown')
# weak model outputs plain prose -> AgentParsingError

# after
agent = CodeAgent(model=stronger_model, tools=[], code_block_tags='markdown')
# or retry loop:
from smolagents.exceptions import AgentParsingError
for _ in range(3):
    try:
        result = agent.run(task); break
    except AgentParsingError:
        continue

Defensive patterns

Strategy: retry

Try / catch

from smolagents.exceptions import AgentParsingError

for attempt in range(3):
    try:
        result = agent.run(task)
        break
    except AgentParsingError as e:
        if 'code parsing' not in str(e):
            raise
        if attempt == 2:
            raise

Prevention

Keep code_block_tags consistent with your prompt template's code fencing instructions.
Use 'markdown' tags with the stock code_agent.yaml templates.
Prefer stronger models or structured outputs for reliable code formatting.

When it happens

Trigger: The model returns prose or malformed code blocks so parse_code_blobs finds no valid code between the configured tags; or with structured outputs, output_text is not valid JSON with a 'code' key.

Common situations: Small/local models ignoring the code-format prompt; code_block_tags mismatch between what the prompt tells the model and what the parser looks for; custom prompt templates that don't instruct ```python fencing when tags are 'markdown'.

Related errors

AI-assisted analysis of huggingface/smolagents@30bb116109 (2026-08-28). Data as JSON: /api/errors/3595bac98adf6e1b. Report an issue: GitHub.