Open-source DeepResearch – Freeing our search agents

OpenAI's Deep Research made a splash by browsing the web, chaining reasoning steps, and answering hard multi-part questions at nearly 67% accuracy on GAIA, a benchmark built specifically to torture-test AI agents. Hugging Face looked at that, shrugged off the mystery around OpenAI's undisclosed agentic framework, and gave themselves 24 hours to build an open equivalent. That's the kind of deadline that either produces something scrappy and forgettable or something genuinely useful. This time it's the latter.

The trick wasn't a smarter model. It was letting the agent write its actions as code instead of the usual JSON blobs. Citing research from Wang et al., the team points out that code lets an agent bundle multiple parallel actions into a single step instead of spelling out twenty separate JSON instructions one at a time. That alone cuts steps by roughly 30%, which translates directly into cheaper runs since LLM calls are the main cost driver. It also means the agent can stash an image or audio clip as a variable and reuse it steps later, rather than hoping the model remembers a dictionary key it invented four turns ago.

The tools themselves are refreshingly unglamorous: a bare-bones text-based web browser and a basic text inspector, both borrowed from Microsoft Research's Magentic-One project with minimal changes. No fancy vision-based browsing yet, no Operator-style mouse-and-keyboard control. Just enough plumbing to prove the concept works. And it does — the setup jumped from Magentic-One's previous open benchmark of about 46% up to 55.15% on GAIA's validation set. Swap the code-writing agent back to plain JSON actions, using the exact same setup otherwise, and performance craters to 33%. That's the whole story in one comparison: the framework, not the underlying model, is doing the heavy lifting.

Hugging Face is upfront that this is nowhere near feature parity with OpenAI's system, particularly on browser interaction, and they're already building toward GUI agents that can see a screen and operate a mouse and keyboard directly, chasing what Operator apparently gives Deep Research today. Other open attempts have popped up too, from developers like dzhng, Assaf Elovic, Nicolas Camara, Jina AI and Matt Shumer, each wiring together their own mix of search, indexing and LLM calls. Hugging Face says it wants to benchmark all of this properly, including swapping in open models like DeepSeek R1, and it's hiring an engineer to keep pushing. The company is also opening a live demo and the full smolagents code, so anyone curious can poke at it themselves rather than take the numbers on faith.