GAIA Level-1 Agent

A tool-using agent for the Hugging Face Agents Course final assignment, built directly on the Anthropic Messages API with no agent framework.

This is a static Space: it publishes the source code and the full run traces so the submission can be verified. The agent runs locally via run_local.py; that is how the submitted answers were produced.

What is here

• gaia_agent/ — the agent loop, tools and answer scoring
• traces/ — one JSON trace per question: every model step, tool call and token count
• answers.json — the submitted answers with confidences
• README.md — how it works and how to run it

Design

• Plain while stop_reason == "tool_use" loop with adaptive thinking and prompt caching.
• Tools: hosted web search and fetch, Wikipedia by dated revision, Python execution, spreadsheet and PDF reading, image inspection, audio transcription, YouTube transcript and frames, chess analysis.
• Hosts that mirror the GAIA dataset are blocked in every search and fetch tool, and any page quoting the question verbatim is discarded, so answers come from primary sources.