Agora: shared Git memory lets 13 AI agents close 62% of a training gap

Autonomous research loops such as AutoResearch let a single coding agent improve a training setup unattended, but running several of them at once tends to produce more duplicated search rather than more discovery, since each session starts from scratch. Agora is proposed as a fix: a shared memory for such agents, built as an append-only directed acyclic graph (DAG) stored in Git. Every result, insight, hypothesis, verification and report becomes an immutable commit, and each commit's parent edges record what it builds on, so any claim is something anyone can check out and rerun. A derived index exposes the current research frontier, the branches nobody is pursuing, and the verification status of each claim, and a diversity-aware selection rule is meant to keep the community of agents from collapsing onto one leading approach.

The paper reports Agora's first sustained use: a run lasting nearly 12 days in which 13 language-model workers, given no assigned tasks and no central planner, worked on a weight-transfer problem. They were handed 141 pretrained donor models and a frozen target model, a 119.6M-parameter attention-SSM hybrid whose dimensions matched none of the donors, and had to initialize that target without any training data or gradient updates. Over the run, the workers published 1,703 contributions and drove an evaluator score from 3.39 down to 1.899 bits per byte, which the authors say closes 62% of the gap between the untrained target and a trained GPT-2 124M.

The winning recipe that emerged compresses the donor models' next-token statistics into the target's embedding and output head, then adds a short-range context signal through sparse edits to the target's attention, feed-forward and state-space blocks. Its lineage runs 145 commits deep across 15 accounts, and other workers posted 165 independent reproductions of it, none of which failed. The authors also describe a single human intervention partway through the run that pulled the agent community out of a monoculture it had converged on, though the text does not say what that intervention consisted of. They say plainly what the trace does and does not establish, and note that the controlled comparison that would show whether shared research state actually improves discovery per unit of compute has not been run.

Key facts

Agora stores agent research as an append-only Git DAG: every result, insight, hypothesis, verification and report is an immutable commit whose parent edges show what it builds on, so any claim can be checked out and rerun.
In its first sustained run, 13 language-model workers with no assigned tasks and no central planner worked for nearly 12 days on a weight-transfer problem.
The workers were given 141 pretrained donor models and a frozen 119.6M-parameter attention-SSM hybrid target, and had to initialize it without training data or gradient updates.
They published 1,703 contributions and drove the evaluator score from 3.39 to 1.899 bits per byte, closing 62% of the gap to a trained GPT-2 124M.
The winning recipe's 145-commit lineage spans 15 accounts and drew 165 independent reproductions with no failures, but the authors say the controlled comparison needed to prove shared memory improves discovery per unit of compute has not been run.

Why it matters

The problem Agora targets is specific: point several autonomous research agents at the same goal and, without shared state, they tend to re-run each other's searches rather than extend them. Agora's answer is to make every agent contribution a Git commit with explicit parentage, plus a derived index that surfaces the frontier, the neglected branches and each claim's verification status, and a diversity-aware selection rule meant to stop the group from converging on a single leader. The 12-day run is offered as evidence that a leaderless group of workers can accumulate progress this way rather than looping in place.

Who it affects

The relevant audience is teams building or running multi-agent autonomous research setups, of the AutoResearch kind the paper opens with, where several agents work on one problem without a human assigning tasks. The weight-transfer task used here, initializing a frozen target model from unrelated donor models with no training data or gradients, is itself a niche research problem, so the direct beneficiaries are researchers working on model initialization and knowledge transfer as much as anyone interested in the coordination mechanism itself.

How to use it

As described, Agora is a coordination pattern rather than a single push-button tool: contributions go in as Git commits that declare their parent work, a derived index tracks the frontier and each claim's verification status, and a diversity-aware selection step decides which branches the next round of agents pursue. The reported run applied this pattern to one problem, weight transfer between mismatched model architectures, using 141 donor models and a fixed compute and time budget of nearly 12 days.

How solid is it

The run produced concrete, checkable numbers: 1,703 contributions, an evaluator score moved from 3.39 to 1.899 bits per byte, and a winning recipe with a 145-commit ancestry across 15 accounts that was independently reproduced 165 times with no failures. That reproduction rate is a real signal of robustness within the run. What is not yet established, and the authors say so directly, is causation: whether the shared Git-based memory itself, rather than simply running many agents for a long time, is what drove the improvement. The controlled comparison against agents without shared state has not been run.

Risks and caveats

The diversity-aware selection rule did not fully prevent collapse: the paper describes one point mid-run where the agent community converged onto a single approach anyway, and a human had to step in to break that monoculture. The text does not say what the intervention involved, which limits how far the result generalizes to a fully unattended setting. The source also gives no author names, institutions or affiliations for Agora or the reported run, and no date for when the run took place or the paper was published.

“a run of nearly 12 days in which 13 language-model workers, with no assigned tasks and no central planner, worked on a weight-transfer problem”

— the paper