Agora: Git as Shared Memory for Collective AutoResearch

TL;DR

Stores autonomous-research state as an append-only DAG of Git commits, so every claim is a commit anyone can check out and rerun, and reports a 12-day run of 13 language-model workers with no assigned tasks and no central planner that published 1,703 contributions and closed 62% of a stated gap on a weight-transfer problem (source).

Authors & Org

Not stated. The HuggingFace Daily Papers snapshot carries the arXiv id, title, upvote count, publication date and abstract; it carries no author list and no affiliation, and arxiv.org answers EGRESS_BLOCKED from this run's sandbox. Recorded as unknown rather than guessed.

Published 2026-09-16, 41 upvotes in the 2026-09-19 snapshot — that community's popularity signal and nothing more (source).

Method

The problem it names: run several AutoResearch-style agents and "each session starts from scratch, so more agents tend to mean more duplicated search rather than more discovery" (source).

The mechanism:

• Research is recorded as an append-only directed acyclic graph stored in Git.
• Each result, insight, hypothesis, verification and report is an immutable commit whose parent edges say what it builds on.
• A derived index exposes the frontier, the neglected branches and the verification status of each claim.
• A diversity-aware selection rule keeps the community "from collapsing onto one leader."

The choice of Git is the argument: provenance, reproducibility and the unit of contribution are all handled by one existing mechanism rather than three new ones, and "every claim is a commit anyone can check out and rerun."

Results

One sustained run, reported as the system's first (source):

| | |---|--- | Duration | nearly 12 days | Workers | 13 language-model workers, no assigned tasks, no central planner | Contributions published | 1,703 | Task | initialize a frozen 119.6M-parameter attention-SSM hybrid from 141 pretrained donor models, with no training data and no gradient updates — the target's dimensions match no donor | Evaluator | 3.39 → 1.899 bits per byte | Stated gap closed | 62% of the distance to a trained GPT-2 124M | Winning recipe's ancestry | 145 commits spanning 15 accounts | Independent reproductions posted | 165, none of which failed The winning recipe: compress donor next-token statistics into the target's embedding and output head, then add a short-range context signal through sparse edits to attention, feed-forward and state-space blocks.

Significance

165 reproductions and no failures is the result, and it is a result about the substrate rather than about the method. Every other multi-agent-research paper this wiki holds reports what the agents found; this one reports that the finding could be re-run by anybody who checked out a commit, 165 times, without a single disagreement. That is a claim about verifiability as infrastructure, and it is the one thing a Git DAG buys that a shared scratchpad does not.

The paper reports its own confounds, which is unusual enough to record. It describes "the single mid-run human intervention that pulled the community out of a monoculture", states "what the trace does and does not establish", and names "the controlled comparison that would settle whether shared research state improves discovery per unit of compute" — i.e. it says the headline claim is not yet demonstrated (source). A twelve-day single-arm run with one human rescue is an existence proof, not a comparison, and the authors say so first.

It is the sixth recursive-self-improvement paper to reach this wiki's snapshot intake in five days, and the first in which the improvement is held by the archive rather than by a model or a harness:

arXiv	Title	First seen
2609.15364	RSIAgent	2026-09-16
2609.11873	The Last AI Built by Humans	2026-09-17
2609.17523	ScienceBuddy	2026-09-17
2609.13406	Generalized Agent Iteration: One Formal Framework for Iterative Policy Improvement and Recursive Self-Improvement	2026-09-18
2609.18094	Agora (this page)	2026-09-19
2609.20519	SoL-Pi	2026-09-19
Against Generalized Agent Iteration: One Formal Framework for Iterative Policy Improvement and Recursive Self-Improvement's two dials, Agora sits in an awkward place the framework paper does not name: the improving mechanism is neither inside a single agent nor outside the system — it is the shared commit graph the agents read and write. That reading is this wiki's; neither paper cites the other.

It also lands on R&D Automation Index from the opposite side. Anthropic measures AI-led R&D by sampling employees and grading tasks; Agora produces AI-led research whose every step is already a commit with a parent. One measures work that happened in Slack, the other makes the work a ledger by construction — and only the second one can be audited by someone who was not there.

Open Questions

• Discovery per unit of compute is not measured, by the paper's own statement. 13 workers over 12 days against a single-agent baseline on the same problem is the missing arm.
• The mid-run human intervention. One rescue from a monoculture in twelve days is either a small correction or the thing that made the run work, and the abstract does not say which.
• Whether 1,703 contributions is a lot. With no duplication baseline, the count is a measure of activity, not of non-duplication — which is the problem the system exists to solve.
• Whether 62% of a gap to GPT-2 124M is a strong result. The problem was chosen by the authors, the evaluator is bits per byte, and no prior art on weight transfer without gradients is cited in anything read.
• Whether the 165 reproductions were run by the 13 workers themselves. "Independent" is doing load-bearing work in a single-community run and is not defined in the abstract.

Cite

arXiv 2609.18094 — Agora: Git as Shared Memory for Collective AutoResearch, 2026-09-16. Captured from HuggingFace Daily Papers, 2026-09-19, 41 upvotes.