# **Open Source “Deep Research” Initiative Demonstrates That Agent Frameworks Enhance AI Model Performance**
## **Overview**
In a decisive effort to contest proprietary AI research tools, Hugging Face has launched an open-source AI research agent known as **”Open Deep Research.”** This project was introduced merely a day after OpenAI revealed its **Deep Research** feature, which autonomously navigates the web and produces research reports. Open Deep Research aims to rival OpenAI’s effectiveness while ensuring the technology remains accessible to developers at no cost.
## **Emergence of AI Research Agents**
AI research agents are crafted to boost the capabilities of large language models (LLMs) by allowing them to execute multi-step tasks. These tasks encompass collecting information from various sources, synthesizing data, and generating organized research reports.
Hugging Face’s Open Deep Research initiative reflects OpenAI’s Deep Research and Google’s Gemini-based Deep Research, both utilizing **agent frameworks** to enhance AI functionality. By introducing an agentic layer to current AI models, these frameworks empower AI to independently undertake complex research activities.
## **Benchmark Comparison: Open Source vs. Proprietary AI**
Despite being developed in just 24 hours, Open Deep Research has already reached **55.15% accuracy** on the **General AI Assistants (GAIA) benchmark**, which assesses an AI model’s capability to collect and synthesize information. In contrast, OpenAI’s Deep Research achieved **67.36% accuracy** on the identical benchmark.
GAIA features intricate multi-step inquiries, such as:
> *”Which of the fruits depicted in the 2008 painting ‘Embroidery from Uzbekistan’ were included in the October 1949 breakfast menu for the ocean liner that subsequently served as a floating set for the film ‘The Last Voyage’? Provide the items as a comma-separated list, arranged in clockwise order according to their positioning in the painting starting from the 12 o’clock point. Use the plural form for each fruit.”*
To accurately respond to such questions, an AI agent must collect and synthesize information from various sources, showcasing the effectiveness of agentic AI in managing complex research operations.
## **Selecting the Appropriate Core AI Model**
An AI agent’s effectiveness is contingent upon the foundational AI model it employs. Open Deep Research is currently based on OpenAI’s **GPT-4o** and simulated reasoning models such as **o1** and **o3-mini** via API. Nevertheless, it remains adaptable to open-weight AI models, providing flexibility for developers.
Hugging Face’s **Aymeric Roucher**, who spearheads the Open Deep Research project, elucidated their model selection criteria:
> *”It’s not ‘open weights’ since we utilized a closed weights model simply because it performed well, but we document the entire development process and make the code available. It can be switched to any other model, allowing for a completely open pipeline.”*
Roucher also experimented with various models, including **Deepseek R1** and **o3-mini**, ultimately determining that **o1** was optimal for this application. Nonetheless, Hugging Face’s **open-R1 initiative** may soon offer an even better open-source alternative.
### **The Significance of Agentic Frameworks**
The primary innovation in Open Deep Research lies in its **agentic design**, which considerably boosts AI performance. Benchmark results underscore this influence:
– **GPT-4o alone (without an agentic framework):** **29% accuracy** on GAIA
– **OpenAI’s Deep Research (with an agentic framework):** **67% accuracy**
– **Hugging Face’s Open Deep Research (with an agentic framework):** **55.15% accuracy**
This highlights that **agent frameworks significantly enhance LLM capabilities**, making them more adept at intricate reasoning and multi-step tasks.
### **The Contribution of “Smolagents”**
A vital aspect of Open Deep Research’s success is Hugging Face’s **”smolagents”** library. This library employs **”code agents”** in place of conventional **JSON-based agents**, rendering AI **30% more effective** in task completion. Code agents execute their actions in programming code, enabling more succinct and efficient task performance.
## **The Velocity of Open Source AI Development**
One of the most notable benefits of open-source AI is the **swift advancement**. Open Deep Research has already underwent significant enhancements, courtesy of the developer community’s contributions. The project also builds upon existing resources, such as **Microsoft Research’s Magnetic-One agent**, which improves web browsing and text examination functionalities.
While Open Deep Research has yet to reach the performance level of OpenAI’s Deep Research, its open-source framework permits developers to **examine, modify, and enhance the technology freely**. This democratization of AI research tools marks a significant progression toward making advanced AI capabilities accessible to everyone.
## **Prospective Developments and Expansions**
Hugging Face is diligently working
