Can decentralized agents coordinate research without a central planner?

Agora's claim is that research agents running in separate, independently scheduled sessions can accumulate progress if their contributions live in a shared, append-only record instead of in transcripts. The abstract describes a directed acyclic graph in Git where each commit records "a result, insight, hypothesis, verification, or report" and links it to prior work. In the reported run, 13 language-model workers "with no assigned tasks or central planner" spent nearly 12 days on a weight-transfer problem: initializing a frozen 119.6M-parameter attention–SSM hybrid from 141 donor models without training data or gradient updates. They published 1,703 contributions, cut the development evaluator score from 3.39 to 1.899 bits per byte, and closed 62% of the gap to a trained GPT-2 124M. The best method's 145-commit ancestry spans 15 accounts, and the record holds 165 reproductions across 95 targets.

The introduction locates the problem in what a lone session leaves behind. If findings stay "in a transcript or a temporary worktree," later sessions must reconstruct "which learning rates diverged, which branches were abandoned, and which results still need independent reproduction." Git supplies persistence and lineage. Searchable views on top show leading results, neglected branches, and verification status, and diversity-aware recommendations point toward experiments beyond the current leaders. Coordination is carried by the artifact rather than by conversation or role assignment, which the introduction sets against multi-agent frameworks that "organize conversations or encode role-specific workflows." The framing is institutional: shared knowledge is what makes problems tractable.

This extends Can decentralized teams outperform central planners in long-running science?. AutoScientists gets decentralization from teams that self-organize around a shared experimental state. Agora adds a different piece, a durable, versioned, verifiable lineage that outlives any single session, and so is a second point of support for dropping the central planner. It also sits at a different place on the record-to-insight spectrum than Can AI research itself without losing human oversight?: the Agora excerpt describes lineage and views but no analyzer stage that distills outcomes. Likewise How can agent systems share learned skills across users? argues for a central evolver, while the Agora excerpt describes no aggregator beyond the graph that workers read directly.

The excerpt reports a single run with no baseline. It does not compare against workers without shared memory, so it cannot say how much of the 62% depends on the record; the authors point to a matched fixed-compute evaluation (Appendix C) as the way to measure that. The score is a development evaluator score, and the excerpt is silent on any held-out result. It also notes "prolonged concentration on one approach": workers refined the bigram recipe for five days, then began exploring state-space edits within a day of the diversity views being deployed. That is a timing observation, not a demonstrated effect of the views. The excerpt does not say how the 15 accounts in the best ancestry relate to the 13 workers. At the strength the evidence allows, the run shows that a persistent, linked record is enough for uncoordinated agents to sustain progress on a hard problem, and that an unmanaged shared record can itself let effort pile onto one line of work. It does not show that shared memory beats other coordination designs.