Agora’s Git-based approach to shared memory for research agents

A post describing an Nvidia paper says Agora records results, hypotheses and verifications as immutable Git commits. Links show what each claim builds on, while an index tracks open branches and verified claims.

The post reports that 13 language-model workers ran for nearly 12 days without assigned tasks or a central planner, making 1,703 contributions. Their task was to initialize a 119.6-million-parameter hybrid model from 141 donor models without training data or gradient updates.

According to the post, the workers reduced the evaluator score from 3.39 to 1.899 bits per byte, closing 62% of the gap to a trained GPT-2 124M. It also says all 165 independent reproductions succeeded.