Agora Turns Git Into Shared Memory for 13 AutoResearch Agents

Thirteen language-model workers ran for nearly twelve days with no assigned tasks and no central planner, and published 1,703 contributions to a shared Git repository. The paper describing Agora frames the mechanism plainly: 'research is recorded as an append-only directed acyclic graph (DAG) stored in Git, so that every claim is a commit anyone can check out and rerun.'

The task the authors handed the workers was concrete. Given 141 pretrained donor models and a frozen 119.6M-parameter attention-SSM hybrid 'whose dimensions match no donor,' the agents had to initialize the target 'without training data or gradient updates.' Over the run, the evaluator moved from 3.39 to 1.899 bits per byte, which the abstract describes as 'closing 62% of the gap to a trained GPT-2 124M.'

The winning recipe was checked, and checked again. 'Its 145-commit ancestry spans 15 accounts, and 165 independent reproductions were posted, none of which failed,' the abstract states. A diversity-aware selection rule, the authors write, 'keeps the community from collapsing onto one leader,' and they describe a 'single mid-run human intervention that pulled the community out of a monoculture' without saying more in the abstract about what that intervention was.

The authors state their own limits. They flag 'the controlled comparison that would settle whether shared research state improves discovery per unit of compute' as work still to come — meaning the run demonstrates that shared-DAG coordination is feasible at 13 agents, not that it beats simply running one agent longer. The paper lands amid a run of tracker coverage on multi-agent research systems; ScienceIDE hit our agents feed the same day.