LangGraph-based Agents for the GAIA Benchmark

Hugging Face Agent Course Final Project

This project is the final hands-on challenge of the Hugging Face Agents Course. The mission: create a sophisticated AI agent capable of tackling real-world tasks from the GAIA benchmark - a rigorous evaluation framework designed to test AI assistants on complex scenarios requiring reasoning, multimodal understanding, web browsing, and advanced tool use.

The Dataset: GAIA

Uses a set of 20 questions selected from the Level 1 validation set of GAIA

Agents

This project implements various types of agents, each with a unique architecture for tackling complex tasks.

1. ReAct-style Agent

A LangGraph-based agent that uses the ReAct (Reason + Act) approach. It makes decisions based on its internal reasoning and chooses tools like web search, Python, or YouTube analysis to gather information before answering.

2. Reflection Agent

An advanced version of the ReAct agent that includes a self-reflection step. After an initial answer, it evaluates its own reasoning and decides whether more tool use is needed before finalizing a response.

3. Plan-Execute Agent

Implements a Plan-Execute architecture using LangChain and LangGraph. The agent generates a step-by-step plan and executes it through a reasoning loop, validating the final answer.

4. Multi-Agent System

Implements a multi-agent architecture using LangGraph and LangChain. A central supervisor agent coordinates a group of specialized agents and dynamically routes tasks to them.

Project Structure

The project is organized into the following modules under src/gaia_agent/:

agents/

This is the core directory containing the implementation of the different agent architectures.

• base.py: Defines a simple Agent base class that manages the interaction with the LangGraph graph, handling state, configuration, and the execution flow.

• react/: Implements a ReAct-style agent.

• reflection/: Implements an agent with a reflection step.

• plan_execute/: Implements a plan-and-execute agent.

• multi_agent/: Implements a multi-agent system with a supervisor.

common/

This directory contains modules with code shared across the different agents.

• tools.py: This is a crucial module that defines all the tools available to the agents. These tools include:

  • excel_tool: For querying Excel files.
  • run_python: For executing Python scripts.
  • audio_model: For processing audio files.
  • visual_model: For processing images.
  • youtube_video_model: For analyzing YouTube videos.
  • web_search: For searching the web.
  • wikipedia_search_html: For searching Wikipedia and getting HTML content.
  • website_scrape: For scraping websites.

• nodes.py: Contains the validate_answer node, which is used by all agents to format the final answer according to the user's request.

• prompts.py: Provides a load_prompt function to load prompt templates from text files.

prompts/

This directory contains the text files for the prompts used by the agents.

• planner.txt: The prompt for the planner in the plan-and-execute agent.
• reflection.txt: The prompt for the reflection step in the reflection agent.
• replanner.txt: The prompt for the replanner in the plan-and-execute agent.
• supervisor.txt: The prompt for the supervisor in the multi-agent system.