Researchers Release Git-Based AI Research Agent Platform 'Agora'… Closes 62% of GPT-2 Gap Without Training

Researchers including Zhang, Yifan released the Agora platform. This platform is designed to allow AI research agents to collaborate and solve complex research tasks without a central manager by using Git repositories as shared memory.

Git-Based Shared Memory Structure

Agora adopts an append-only DAG structure that records all research results and hypotheses as immutable commits. The Git history becomes the sole state, and agents only read from and write to it. Verified evidence propagates to parent node scores, balancing exploration and exploitation through UCB(Upper Confidence Bound) rules.

Experimental Results: Training-Free Model Initialization

In experiments, agents successfully initialized a frozen 119.6M hybrid model to 1.899 bpb without training data or gradient updates. This significantly improved performance compared to random initialization (3.3923 bpb), closing 62% of the performance gap with the trained GPT-2 124M model (approx. 1.0 bpb). The winning recipe led to 145 commits across 15 accounts, and all 165 independent reproduction attempts succeeded.

Collaboration Dynamics

Scored contributions on the first day accounted for approximately 98% of the total performance reduction, demonstrating the importance of initial exploration. A monoculture phenomenon occurred for specific recipes, but it was overcome through human intervention.