What causes "Error while parsing tool call from model output: {e}" in huggingface/smolagents?

Switch to a stronger or tool-calling-native model (the model returns proper tool_calls, bypassing parse_tool_calls).

huggingface/smolagents · error · AgentParsingError

Error while parsing tool call from model output: {e}

Error message

Error while parsing tool call from model output: {e}

What it means

When the model returns a message without native tool_calls, ToolCallingAgent._step_stream calls model.parse_tool_calls() to convert the text into tool calls; any exception there is wrapped in AgentParsingError. This usually means the model produced output that doesn't conform to the expected tool-call format (bad JSON arguments, malformed function-call syntax).

Source

Thrown at src/smolagents/agents.py:1331

                )
                self.logger.log_markdown(
                    content=str(chat_message.content or chat_message.raw or ""),
                    title="Output message of the LLM:",
                    level=LogLevel.DEBUG,
                )

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

        if chat_message.tool_calls is None or len(chat_message.tool_calls) == 0:
            try:
                chat_message = self.model.parse_tool_calls(chat_message)
            except Exception as e:
                raise AgentParsingError(f"Error while parsing tool call from model output: {e}", self.logger)
        else:
            for tool_call in chat_message.tool_calls:
                tool_call.function.arguments = parse_json_if_needed(tool_call.function.arguments)
        final_answer, got_final_answer = None, False
        for output in self.process_tool_calls(chat_message, memory_step):
            yield output
            if isinstance(output, ToolOutput):
                if output.is_final_answer:
                    if len(chat_message.tool_calls) > 1:
                        raise AgentExecutionError(
                            "If you want to return an answer, please do not perform any other tool calls than the final answer tool call!",
                            self.logger,
                        )
                    if got_final_answer:
                        raise AgentToolExecutionError(
                            "You returned multiple final answers. Please return only one single final answer!",
                            self.logger,
                        )

View on GitHub (pinned to 30bb116109)

Solutions

Switch to a stronger or tool-calling-native model (the model returns proper tool_calls, bypassing parse_tool_calls).
Retry the run — a resampled completion often parses cleanly; add retry-on-AgentParsingError logic.
Upgrade smolagents to pick up parser fixes for common malformed-JSON patterns.

Example fix

from smolagents.exceptions import AgentParsingError

try:
    result = agent.run(task)
except AgentParsingError:
    result = agent.run(task)  # resample

Defensive patterns

Strategy: retry

Try / catch

from smolagents.exceptions import AgentParsingError

for attempt in range(3):
    try:
        result = agent.run(task)
        break
    except AgentParsingError:
        if attempt == 2:
            raise

Prevention

Prefer models with native tool calling to avoid text-based tool-call parsing.
Keep tool argument schemas simple (flat JSON with primitive types).
Retain parsing errors in memory so the model self-corrects on retry.

When it happens

Trigger: agent.run()/step() where the provider returns tool calls as text that fails parsing — e.g. JSON arguments with single quotes/trailing commas, or a weaker model that free-forms the tool-call schema.

Common situations: Using small/local models that emit slightly malformed JSON; provider API changes altering the tool-call payload shape; prompt templates that don't constrain output format.

Related errors

AI-assisted analysis of huggingface/smolagents@30bb116109 (2026-08-28). Data as JSON: /api/errors/19b937998ec3b40a. Report an issue: GitHub.