Hugging Face tries to clone OpenAI’s Deep Research in 24 hours with open code agent

OpenAI’s new Deep Research tool is undeniably slick—it browses the web, synthesizes multi-step answers, and just scored a mind-bending 67% on the GAIA benchmark’s validation set, a massive leap from GPT-4’s paltry sub-7% solo performance. But its inner workings are a black box. That proprietary “agentic framework” is the real secret sauce, and OpenAI isn’t sharing the recipe. So a team at Hugging Face did what open-source devs do best: they set a 24-hour timer to build a public replica and reverse-engineer those gains.

The linchpin of their approach is something called a “code agent.” The core insight, backed by research from Wang et al., is that making a model write its action plan in actual Python code is far more precise for complex, multi-step tasks than typical JSON-based tool-calling. Code is purpose-built for expressing intricate sequences, error handling, and logic. When paired with a powerful free model like DeepSeek R1, the team argues this setup can close the chasm between raw model performance and agentic superpowers—sometimes adding up to 60 percentage points on certain benchmarks.

GAIA, the benchmark in question, is a brutal test. It demands an agent correctly chain together disparate information, like identifying fruit in a specific painting, cross-referencing that with a 1940s ocean liner’s breakfast menu, and outputting it all in a strict clockwise order. This is a torture test for planning and rigorous tool use, two areas where a standalone LLM falls flat on its face. Hugging Face is betting that by stripping out proprietary black-box magic and replacing it with transparent, code-driven loops, they can democratize a research capability that OpenAI is currently charging $200 a month for.

Full parity, they readily admit, is a longer-term project. OpenAI’s system also leans on advanced browser GUIs—think Operator—that go beyond the text-based web interactions in this first open attempt. But even a partial clone would be a big deal. It puts a powerful research assistant into the hands of anyone with a GPU, and more importantly, lets the community pick apart and improve the agentic logic that OpenAI is keeping under lock and key. The race to open-source the reasoning behind the reasoning is officially on.