GAIA Benchmark Agent Evaluation

Implementation Overview:

• Orchestration: ReAct-style agent built with LangGraph (StateGraph + ToolNode) and Gemini (langchain-google-genai) with built-in rate limiting (15 RPM).
• Tools:
  • web_search & extract_webpage: Web search and full-page text extraction powered by Tavily.
  • python_interpreter: Subprocess Python execution for calculations, logic puzzles, .py scripts, and .xlsx spreadsheets (pandas/openpyxl).
  • analyze_media: Multimodal inspection of YouTube videos, .mp3 audio, and .png images via the google-genai SDK.

Log in with your Hugging Face account below and click Run Evaluation & Submit All Answers to evaluate the agent on the GAIA benchmark subset.