Why is memory_step.action_output not assigned when tool_name != "final_answer" in ToolCallingAgent.step()?

[iamxtso]

In the ToolCallingAgent.step() method, when tool_name == "final_answer", memory_step.action_output is assigned the value of final_answer and is returned. However, when tool_name != "final_answer", memory_step.action_output is not assigned a value.

class ToolCallingAgent(MultiStepAgent):
    def step(self, memory_step: ActionStep) -> Union[None, Any]:
        if tool_name == "final_answer":
            # ✅ action_output is assigned.
            memory_step.action_output = final_answer  
            return final_answer  
        else:
            # ❓ Why is action_output not assigned?
            memory_step.observations = updated_information  
            return None

[potatohoney-p]

In the code you quoted, return None means “this step is not the final answer; continue the agent loop.” It does not mean that the tool result is discarded: the observation is the representation sent back to the model. action_output should not be treated as a general-purpose archive of every raw tool return.

The implementation has changed since that snippet. On current main, a ToolCallingAgent can execute multiple calls in one step. Each result is emitted separately as a ToolOutput, while ActionStep.observations stores the combined textual observations. If you need the actual Python values rather than their string representation, consume those events:

from smolagents.agents import ToolOutput

outputs_by_call_id = {}
for event in agent.run(task, stream=True):
    if isinstance(event, ToolOutput) and not event.is_final_answer:
        outputs_by_call_id[event.id] = event.output

Fully iterate the generator so the run completes. This is run(stream=True), not a requirement to enable token streaming with stream_outputs=True on the model/agent.

If you only need the readable trace after a run:

from smolagents.memory import ActionStep

for step in agent.memory.steps:
    if isinstance(step, ActionStep):
        print(step.step_number, step.observations)

The distinction is visible in process_tool_calls: ToolOutput.output receives the raw result, and observations are accumulated separately. ActionStep.to_messages then includes those observations in subsequent model input. Thus an unset action_output alone is not evidence of lost results. The event example is for the current API; an older installed version may still have the step() implementation from your question.