by Hugging Face
Open Deep Research is Hugging Face's open-source reproduction of a web-browsing research agent, built in a 24-hour effort. It uses an agent framework that writes its actions as code.
AgentFree
Step 1: Read the Hugging Face blog post to see how the agent is built.
Step 2: Open the linked GitHub code for the project.
Step 3: Choose a language model to drive the agent.
Step 4: Run the agent on a research question or a GAIA task.
Step 5: Compare results with the benchmark figures in the post.
The post says code actions are more concise than JSON and need about 30% fewer steps, which cuts the tokens generated.
GAIA, a benchmark of hard multi-step questions that involve tools, planning and constrained answer formats.
No. The authors call it a work in progress and say parity needs better browser use beyond text-only web interaction.
A layer on top of a language model that lets it take actions like browsing or reading PDFs and organizes them into steps.
BioMCPBioMCP is a single-binary command-line tool and MCP server for querying biomedical databases. It gives researchers, clinicians and agents one command grammar instead of many sites and APIs.
Devv AIAI developer search and coding assistant for technical research, code questions, repository context, and implementation guidance
AviaryAviary is an open-source gymnasium from FutureHouse for defining custom language-agent reinforcement learning environments. It ships with environments for math, general knowledge, biological sequences and scientific literature search.