Agora's Git-based research DAG is plausible but still lacks a controlled baseline

The core claim here is architectural rather than algorithmic: when autonomous research agents share nothing but a transcript, adding more of them produces more duplicated search, not more discovery. Agora's answer is to treat the research record itself as the coordination layer — every result, hypothesis, verification, and negative finding committed to a Git DAG, immutable and content-addressed, so any worker can check out any ancestor and rerun it exactly.

The design choices follow from this premise. Evidence scores are computed from downstream use by other accounts, not from votes or self-citation, so a worker cannot manufacture credibility by extending its own branch. The analyze view deliberately surfaces neglected clusters alongside the leaderboard, because a pure exploitation signal pulls every worker toward the same parent and makes a saturated basin look like progress. A diversity-aware selection rule — a UCB variant that penalizes near-duplicate descriptions and rewards thin clusters — is the mechanism for keeping the community from collapsing.

The 12-day weight-transfer run is the paper's evidence base, and it is worth reading carefully rather than skimming for the headline number. Thirteen agent sessions, given only a two-page brief and no assigned roles, published 1,703 contributions on a problem with no obvious solution path: initialize a frozen 119.6M-parameter attention-SSM hybrid from 141 donor models, no training data, no gradient updates. The winning method compresses donor next-token statistics into a low-rank factorization of the target's embedding and output head, then re-enables sublayers through sparse deterministic edits. Its 145-commit ancestry crosses 15 accounts; 115 of those parent edges cross account boundaries. The community moved from 3.39 to 1.899 bits per byte, closing roughly 62% of the gap to a trained GPT-2 124M.

The coordination dynamics are as instructive as the result. The first 18 scored contributions account for about 98% of the total reduction. After that, the graph shows a narrow spine of successive leaders surrounded by short abandoned branches, and 696 pairs of different accounts posted identical scores — many within an hour of each other. Shared memory did not prevent parallel rediscovery. For five days the community refined one recipe by fractions of a thousandth of a bit per byte per step, each worker reading the same leaderboard. It left that basin within a day of being shown a diversity map — a single human intervention that deployed the clustering and landscape views.

The authors are direct about what the run does not settle. There is no matched comparison: the same models and compute were never run without Agora or with a plain leaderboard. The causal question — does a research DAG improve discovery per unit of compute, or does it just file parallel waste more neatly — is left explicitly open, with a preregisterable evaluation matrix in the appendix. That honesty is the paper's most useful feature for anyone thinking about building on this. The mechanism is plausible and the trace is detailed, but the controlled experiment has not been run yet.