Agora paper: 13 LLM research agents self-organize via Git for 12 days

suchenzang · x · 2026-09-22

The arXiv paper Agora: Git as Shared Memory for Collective AutoResearch stores contributions from multiple research agents as an append-only DAG in Git: each commit records a result, insight, hypothesis, verification or report, linked to prior work. Searchable views surface leading results, neglected branches and verification status, with diversity-aware recommendations for under-explored experiments.

The authors ran a nearly 12-day experiment: 13 LLM workers, with no task assignments or central planner, used Agora to tackle a weight-transfer problem — initializing a frozen 119.6M-parameter attention-SSM hybrid whose dimensions match none of 141 pretrained donor models, with no training data or gradient updates. The run produced 1,703 contributions and improved the evaluator score from 3.39 to 1.899 bits per byte, closing 62% of the gap to a trained GPT-2 124M. The best method compresses donor next-token statistics into the target's embedding and output head, then adds short-range context via sparse edits to attention, FFN and SSM blocks; its 145-commit ancestry spans 15 accounts. Participants also posted 165 independent reproductions across 95 targets with no reported failures.

Original post →

More from coding & agent

coding & agent channel →