app.py

import gradio as gr
from agent import GaiaAgent
from run import run_and_submit_all  # Adjusted if you moved out of `gaia_benchmark`

agent = GaiaAgent()

with gr.Blocks() as demo:
    gr.Markdown("# GAIA Agent Evaluation")
    btn = gr.Button("Run Evaluation & Submit All Answers")
    result = gr.Textbox(label="Submission Result")

    btn.click(fn=lambda: run_and_submit_all(agent), outputs=result)

demo.launch()