GAIA Agent — Final Assignment

This Space documents my agent for the Hugging Face Agents Course final assignment (GAIA benchmark, Unit 4). All answers submitted to the scoring API were produced live by this agent code running on my local machine (Python 3.11), using real tool calls — web search, webpage reading, Python execution, Whisper audio transcription and vision — with no hardcoded or looked-up answers.

Why local instead of inside this Space? Free-tier accounts cannot host Gradio/Docker Spaces in 2026 (402 PRO-subscription wall), and the course's official file-serving endpoint (/files/<task_id>) is broken server-side. The agent therefore ran outside a Space, and this free static Space publishes the exact code and run artifacts for full review.

Architecture

LLM: Groq (openai/gpt-oss-120b for reasoning/search, qwen/qwen3.8-27b for vision + JSON-safe tool calls, whisper-large-v3-turbo for audio), with Google Gemini (gemini-3.8-flash etc.) as a cross-checked second provider.
Tools: web_search (DuckDuckGo), visit_webpage (requests + BeautifulSoup), run_python (subprocess with pandas/openpyxl/chess).
Files: downloaded from the official gated gaia-benchmark/GAIA dataset (validation split) with an HF token; images passed as vision input, audio transcribed via Whisper, Excel parsed with pandas.
Loop: function-calling agent, max 8 steps, forced final answer after budget exhausted; exact-match answer formatting rules in the system prompt.

Files

groq_agent.py — the submitted agent (Groq/OpenAI-compatible brain)
gaia_agent.py — the same agent architecture on Google Gemini (used to cross-check answers)
run_local.py — evaluation harness with caching
answers_groq.json / answers.json — the agents' full run artifacts (answers + per-question timing)

Sample trace (Q6 — non-commutativity counter-examples)

step: run_python({...})  → brute-force check of all ordered pairs against the table
whisper:  "In a saucepan, combine ripe strawberries, granulated sugar, freshly squeezed
          lemon juice and cornstarch..."   (Q10, audio transcription)
vision:   qwen3.8-27b reads the chess board image (Q4)

Submitted via POST /submit of the official scoring API, username summerdesu. Score and full history are visible on the course leaderboard.