Nvidia unveils Agora: shared memory for self-coordinating research agents

Nvidia researchers have posted a preprint on ArXiv describing Agora, a shared-memory system that lets coding agents collaborate on the same research problem without repeating experiments that have already been run. Every result, hypothesis and validation is recorded as an immutable Git commit; parent edges point to the claims each new entry depends on, and a central index tracks open branches and which claims have been verified. The authors say the structure removes the need for a central planner or any predefined task allocation.

In the main experiment the team ran 13 LLM workers for nearly 12 continuous days with no assigned tasks and no central planner. The goal: initialize a 119.6-million-parameter hybrid model from 141 donor models, using no training data and no gradient updates. The agents had to discover for themselves how to combine the weights, validate the outcomes and log everything in Agora. Over the period the agents logged 1,703 contributions.

The primary metric was the evaluator's bits per byte, which fell from 3.39 to 1.899 — closing 62% of the gap to a trained 124-million-parameter GPT-2. Equally important, all 165 independent reproductions succeeded, indicating process stability and confirming that the Git-based logging enables full verification. The paper stresses that the improvement came without retraining, solely through collaborative assembly and validation.

As a preprint that has not yet undergone peer review, the results await external confirmation. The paper does not disclose the total compute cost of running 13 agents for 12 days, nor does it compare the approach with existing model-merging or evolutionary-search methods that rely on a central planner. It also remains unclear how well the methodology generalizes beyond the specific task of initializing a hybrid model. Code and data are available for interactive exploration at academy.dair.ai/papers/agora-g…