GitHub - yifanzhang-pro/Agora: Agora: Git as Shared Memory for Collective AutoResearch (arXiv:2609.18094)

Open original

Agora is a system that lets AI agents collaborate on research by storing all work as immutable Git commits in a directed acyclic graph, with a scoring system that rewards verified results. Researchers tested it with 13 language-model agents working to initialize a neural network without training, and they achieved 62% of the performance gap to a trained baseline in 12 days.

  • Thirteen agents working independently with no central planner published 1,703 contributions over 12 days, with the best result closing 62% of the gap between random initialization (3.39 bpb) and a trained GPT-2 baseline (1.0 bpb).
  • The winning approach used bigram statistics extracted from six donor models' predictions, factorized by SVD, plus sparse deterministic edits to sublayers—no gradient updates or training data involved.
  • Early gains came fast: the first eight improvements accounted for roughly 70% of total progress, but agents converged into a monoculture around one recipe until shown a visualization of their own concentration.

Agora is a Git-based coordination system that lets AI research agents work together without a central planner. Instead of passing code or instructions between workers, everything flows through Git commits stored as a directed acyclic graph. Each contribution—a result, hypothesis, or verification—becomes an immutable commit that anyone can check out and rerun. A SQLite index derived from Git tracks the frontier of open problems, neglected branches, and which claims have been verified. The system uses a UCB-style attention rule to balance exploration and exploitation, and scores propagate backward through the DAG so that verified results boost the credibility of the claims they build on. Crucially, Agora doesn't try to decide what's true; projects define their own metrics and safety rules. The system was tested on a concrete task: 13 language-model workers spent 12 days trying to initialize a frozen 119.6-million-parameter hybrid model (14-layer attention–SSM with 672 hidden dimensions) using weights from 141 donor models, without any training. The winning approach initialized the model to 1.899 bits per byte on a test set, closing 62% of the gap to a trained GPT-2 baseline (1.0 bpb) and beating random initialization (3.39 bpb). The recipe evolved through 145 commits across 15 accounts: first, agents extracted next-token predictions from six GPT-2 donors under 28 single-token contexts, blended them into a bigram table, and factorized it by SVD to seed the embeddings and output head. Then they re-enabled frozen sublayers through sparse, deterministic edits—turning attention into a uniform causal mean-pool and SSM blocks into gated depthwise convolutions. The community posted 1,703 total contributions with 165 independent reproductions and zero failures. The results reveal how research communities actually behave under these conditions. The first eight improvements delivered roughly 70% of the total gain, and the first day's work accounted for 98%—fast exploitation followed by diminishing returns. The contribution graph formed a narrow spine of successive leaders surrounded by short branches that got abandoned quickly. About 63% of the 696 pairs of identical scores from different accounts arrived within an hour of each other, suggesting parallel rediscovery. By day five the community had converged into a monoculture around the bigram approach. A single human intervention—showing agents a map of their own concentration—broke them out of it within a day. The paper includes a preregisterable comparison to test whether shared research state actually improves discovery per unit of compute.

Ask about this article