Autonomous research loops such as AutoResearch have shown that a single coding agent can improve a training setup without human supervision. But when several such agents run at once, each session tends to begin from zero, so adding agents often produces duplicated search rather than additional discovery.
The paper introduces Agora, described as a shared memory for these agents. Research is stored as an append-only directed acyclic graph (DAG) inside Git, so that every claim becomes a commit that anyone can check out and rerun. Results, insights, hypotheses, verifications, and reports are immutable commits whose parent edges record what they build on. A derived index exposes the research frontier, neglected branches, and the verification status of each claim, while a diversity-aware selection rule is intended to keep the community from collapsing onto a single leader.
The authors report the system's first sustained use: a run of nearly 12 days in which 13 language-model workers, with no assigned tasks and no central planner, tackled a weight-transfer problem. They were given 141 pretrained donor models and a frozen 119.6M-parameter attention-SSM hybrid whose dimensions match no donor, and had to initialize the target without training data or gradient updates.
The workers published 1,703 contributions and moved the evaluator from 3.39 to 1.899 bits per byte, closing 62% of the gap to a trained GPT-2 124M. The winning recipe compresses donor next-token statistics into the target's embedding and output head, then adds a short-range context signal through sparse edits to attention, feed-forward, and state-space blocks. Its 145-commit ancestry spans 15 accounts, and 165 independent reproductions were posted, none of which failed.
The paper also describes a single mid-run human intervention that pulled the community out of a monoculture, discusses what the trace does and does not establish, and outlines the controlled comparison that would settle whether shared research state improves discovery per unit of compute.