A preprint posted to arXiv on September 16 (2609.18094, from researchers at NVIDIA and several institutions) proposes institutionalizing multi-agent research collaboration: every experiment, hypothesis, verification and report becomes an immutable commit in a Git repository, so that language-model workers with no assigned tasks and no central planner can share memory and reproduce each other's work like a research community. The authors call the system Agora, after the ancient Greek public square.
The motivation i
[1][2]s blunt. Existing automated research loops such as AutoResearch show that a single coding agent can improve a training setup unattended; but when several such agents run in parallel, each session starts from scratch and none knows that another worker already tested an idea or hit a wall. More agents tend to mean more duplicated search rather than more discovery. Agora reframes this as an institutional-memory problem: not making one agent smarter, but making the work of a population of agents visible and reusable.
Mechanically, Agora models the project as a directed acyclic graph. Every contribution — code, metadata, parent edges, optional evaluation results — corresponds to a Git commit that anyone can check out and rerun. A derived index exposes the frontier, the neglected branches and the verification status of each claim, and a diversity-aware selection rule keeps the community from collapsing onto a single leader.
The paper reports a first sustained run: 13 language-model workers, for nearly 12 days, with no assigned tasks, worked together on a weight-transfer problem — given 141 pretrained donor models and a frozen 119.6M-parameter attention-SSM hybrid whose dimensions match no donor, initialize the target without training data or gradient updates. The workers published 1,703 contributions and drove the evaluation 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 lineage spans 145 commits across 15 accounts, and 165 independent reproductions were posted — none failed.
The paper also records a telling episode: a single mid-run human intervention pulled the community out of a monoculture, meaning the diversity rule does not always prevent every worker from crowding into the same local optimum. And the authors are explicit about the boundary: the run demonstrates that shared-DAG coordination is feasible at 13 agents, but there is no compute-matched control comparing it against simply running one agent longer. The 62% figure is an existence proof, not a productivity claim.
The most striking design choice is the substrate itself: Git, a version-control primitive built for human collaboration, repurposed as the coordination protocol for machine agents. Agents need no dialogue and no shared session context — only a history anyone can inspect and rerun. For a field where multi-agent practice is still mostly "collaboration as conversation," that is a genuinely different default. Whether it is actually faster than centralized or single-agent alternatives will have to wait for the controlled comparison the authors promise. Until then, it offers a clean way of posing the question: where should a research community's memory live?
[1][2]