Agora: Git as Shared Memory for Collective AutoResearch

Summary

Agora (NVIDIA) gives research agents working in separate sessions a shared memory: an append-only directed acyclic graph stored in Git, where each commit records a result, insight, hypothesis, verification or report and links to prior work. Searchable views surface leading results, neglected branches and verification status, and diversity-aware recommendations push experiments away from the current leaders. The reported run is unusually concrete: nearly 12 days, 13 language-model workers, no assigned tasks and no central planner, attacking a weight-transfer problem — initialise a frozen 119.6M attention-SSM hybrid from 141 pretrained donor models, with no training data and no gradient updates. The workers published 1,703 contributions and cut the development evaluator score from 3.39 to 1.899 bits per byte, closing 62% of the gap to a trained GPT-2 124M. The winning method's ancestry spans 145 commits across 15 accounts, and participants posted 165 independent reproductions.

Contribution

Claimed and evidenced: a coordination substrate for many agents working asynchronously without a planner, plus a run showing cumulative progress attributable to it. The provenance numbers (145-commit ancestry across 15 accounts, 165 reproductions) are the strongest part — they demonstrate that contributions actually built on each other rather than running in parallel.

Method

System paper with a single deep deployment. Evaluation is by the scientific outcome (bits per byte on a held-out evaluator) and by process statistics from the Git DAG. There is no comparison against a centrally planned baseline or a single-agent control, so the causal contribution of Agora itself versus 13 workers and 12 days of compute is not isolated.

Relevance to RISE

Directly relevant to the catalog's end-to-end and multi-agent entries, and closest in spirit to deepscientist (a Git repo per research quest) — but Agora makes the repository the coordination mechanism rather than a record. For the RISE pipeline anatomy it is the clearest worked example of persistent shared memory as infrastructure rather than as an agent's private scratchpad. For the Architecture as Epistemology question it is a natural experiment at the extreme of the structure axis: no hierarchy, no planner, pure stigmergic coordination through an append-only artefact, with diversity-aware recommendation as an explicit anti-herding mechanism — which is exactly the intervention the diversity-collapse literature (chen2026diversity, kong2026multillm) says is needed.

Critique / open questions

n=1 deployment on a task with a cheap machine-checkable score (bits per byte); whether the design survives tasks where no such oracle exists is untested, and that regime is where most research lives. No ablation of the diversity-aware recommender, so its contribution to avoiding premature convergence is asserted rather than shown. Cost is not reported in the abstract.

Key quotes

"Agora stores their contributions as an append-only directed acyclic graph (DAG) in Git. Each commit records a result, insight, hypothesis, verification, or report and links it to prior work." (abstract)

"We report a run of nearly 12 days in which 13 language-model workers, with no assigned tasks or central planner, used Agora to solve a weight-transfer problem." (abstract)

"They published 1,703 contributions and reduced the development evaluator score from 3.39 to 1.899 bits per byte, closing 62% of the gap to a trained GPT-2 124M." (abstract)