Agora: Git as Shared Memory for Collective AutoResearch

Agora. Git as Shared Memory for Collective AutoResearch. By Yifan Zhang and seven others, from NVIDIA. Published on the sixteenth of September. Its code repository is a placeholder: a readme, and nothing to run yet.

Autonomous research loops, AutoResearch among them, show that a single coding agent can improve a training setup without anyone watching. The natural next step is to run several at once. But each session starts from scratch, so more agents tend to mean more duplicated search rather than more discovery. Many agents working alone mostly rediscover the same things. The missing piece is not a better agent. It is a way for agents to build on each other's work.

Agora is a shared memory for such agents, and it is built on Git. Research is recorded as a directed acyclic graph that only ever grows, a DAG, stored in a Git repository. Every claim is a commit that anyone can check out and rerun. Each result, insight, hypothesis, verification and report is an immutable commit, and its parent edges say what it builds on. Using Git means the record is not a summary of the research but the research itself. Anyone, human or agent, can go to any claim, see what it came from, and run it again.

On top of that graph, a derived index shows the frontier of current work, the branches being neglected, and whether each claim has been verified. A selection rule that is aware of diversity keeps the community of agents from all collapsing onto whichever line of work is leading. The design gives the agents what a research community needs: a common record, a way to see what is open, and a push to spread out.

The first sustained use ran for nearly twelve days. Thirteen language-model workers, with no assigned tasks and no central planner, took on a weight-transfer problem. They were given one hundred and forty-one pretrained donor models and a frozen target: an attention and state-space hybrid with one hundred and nineteen point six million parameters, whose dimensions matched none of the donors. The task was to initialize that target without any training data and without gradient updates. Everything the workers produced went into the shared graph, so each of those contributions is a commit that can be traced and rerun.

The workers published one thousand seven hundred and three contributions. They drove the evaluator from three point three nine bits per byte down to one point eight nine nine, closing sixty-two percent of the gap to a trained GPT-2 of one hundred and twenty-four million parameters. That ancestry is the point of the system made visible. The best result was not the work of one agent, but a chain of commits built on each other by many.

The winning recipe compresses the donors' next-token statistics into the target's embedding and output head, then adds a short-range context signal through sparse edits to its attention, feed-forward, and state-space blocks. Its ancestry runs to one hundred and forty-five commits across fifteen accounts. One hundred and sixty-five independent reproductions were posted, and none of them failed.

The authors are explicit about the limits of a single trace. They describe the one human intervention, midway through the run, that pulled the community out of a monoculture. They set out what the trace does and does not establish. And they describe the controlled comparison that would settle the real question: whether shared research state improves discovery per unit of compute. What the run does show is that a plain version-control graph can hold a whole community's research, with every step checkable and every result reproducible by anyone who checks it out.