A team led by Yifan Zhang, with contributors including NVIDIA’s Jan Kautz, released Agora, an open-source system that lets autonomous research agents coordinate by writing to a shared Git repository instead of exchanging messages in a chat thread. Every hypothesis, result, and verification becomes a permanent, checkable commit. The project ships as a GitHub repository alongside a paper posted to arXiv on September 16.
Most multi-agent research setups today route coordination through a scratchpad or a running conversation log that gets summarized and forgotten. Agora’s bet is that giving agents a permanent, append-only ledger, the kind version control already gives software teams, produces a durable public record of what worked, what failed, and who built on whose idea. A derived index tracks which claims sit at the frontier and which branches nobody followed up on, and a scoring rule pushes credit back to earlier commits when later work confirms them, while excluding an agent from citing itself.
The team ran a demonstration to test the idea: initializing a brand-new, untrained neural network using only the weights of 141 unrelated open-source models, with no gradient updates and no training data allowed. Thirteen language-model agents worked the problem for close to twelve days with no assigned tasks and no manager coordinating them, and posted 1,703 contributions to the shared graph. Their best solution got the untrained model to a bits-per-byte score of 1.899, versus 3.3923 for a random starting point, closing 62% of the distance to a fully trained reference model of similar size. The underlying reasoning: instead of copying donor model parameters directly, the winning approach queried donor models for their predictions and built the new model’s starting weights from those predictions.
The coordination pattern the paper describes is worth watching independent of the specific benchmark. Eighteen contributions posted on the first day accounted for roughly 98% of the total gain, and the remaining 1,106 contributions spent the next eleven days chasing the final 0.03 points on the bits-per-byte scale. Separate agents landed on identical scores within an hour of each other 63% of the time, evidence that the shared ledger let ideas spread fast rather than each agent working in isolation. By day five, repeated refinement of one approach had crowded out every alternative branch. A one-time human intervention, a visualization showing how concentrated the effort had become, broke that pattern within a day.
That intervention is also the honest caveat here. Agora’s authors are the ones reporting these results, on their own repository, without a published control group that ran the same task without shared memory. The paper itself says the matched comparison needed to prove shared research state actually beats agents working independently has not been run yet. Teams building multi-agent research or coding pipelines should treat the coordination structure, not the specific benchmark numbers, as the transferable idea, and watch for that head-to-head comparison before assuming a Git-backed memory layer beats simpler alternatives.
Based on the Agora project repository and paper published by Yifan Zhang and collaborators on GitHub, September 16, 2026.
