Richer GradioUI for debugging and vsualization

[vitalyshalumov]

Throughout the agent's run, I print several key remarks that I want to display in the Gradio UI. How can I connect them to the Gradio UI?

[potatohoney-p]

In smolagents==1.26.0, GradioUI renders ActionStep.observations as Execution Logs. It does not capture arbitrary server-side stdout.

That means:

• A print(...) in the Python code executed by CodeAgent is already included in the step's execution logs.
• A print(...) inside your own Python tool/callback goes to server stdout, not automatically to the chat UI.

For your own remarks, one way to use the existing UI is to collect them and append them to the completed action's observations:

from smolagents import CodeAgent, GradioUI
from smolagents.memory import ActionStep

remarks = []  # Dedicated to this agent/session, not shared across users.

def publish_remarks(step: ActionStep, agent):
    if remarks:
        step.observations = (
            (step.observations or "") + "\n" + "\n".join(remarks)
        )
        remarks.clear()

# In your own tool, replace print("...") with remarks.append("...").
agent = CodeAgent(
    tools=tools,
    model=model,
    step_callbacks={ActionStep: publish_remarks},
)
GradioUI(agent).launch(share=False)

This displays the remarks when the action step finishes, not continuously during a long-running tool. Also, observations are agent memory: the remarks will be included in later model requests. For UI-only diagnostics, keep them outside observations and render them in a separate Gradio output/event stream instead. Avoid global stdout redirection in a multi-user app, because unrelated requests can share the same stdout.

Verified with Python 3.12.14, smolagents 1.26.0, and Gradio 5.49.1 using a deterministic model stub and the actual GradioUI response generator: code-level print appeared in Execution Logs; host-side print did not; the callback remark appeared both in the Gradio message and the next model request. No live LLM or browser-rendering test was involved.

Sources: Gradio observation rendering, step callback API.