Agora logs 13 agents' research in Git and closes 62% of a weight gap

single source· 1 articles · confidence: medium · first seen 2026-09-15 20:00 UTC

What this means for you

Nothing to act on yet. This is a preprint reporting a single 12-day run; the sources give no code release, no cost figures, and no controlled comparison against agents working separately, which the authors say is what would settle the claim. The idea to watch is the shared append-only log, not the score.

Agora records each agent's results, hypotheses and verifications as append-only commits in Git (a permanent, branched record), so parallel research sessions do not repeat each other. In a roughly 12-day run, 13 language-model workers with no assigned tasks had to initialise a frozen 119.6M-parameter model from 141 pretrained donors, with no training data or gradient updates. They posted 1,703 contributions and cut the evaluator from 3.39 to 1.899 bits per byte, a lower-is-better score for predicting text — 62% of the gap to a trained GPT-2 124M. One human intervened mid-run when the agents converged, and the authors say a controlled comparison is still needed.

Key facts

  • ·13 language-model workers ran for nearly 12 days with no assigned tasks and no central planner. source
  • ·The workers published 1,703 contributions; the winning recipe's ancestry spans 145 commits across 15 accounts. source
  • ·The evaluator moved from 3.39 to 1.899 bits per byte, closing 62% of the gap to a trained GPT-2 124M. source
  • ·165 independent reproductions were posted, none of which failed. source
  • ·The target was a frozen 119.6M-parameter attention-SSM hybrid, initialised from 141 pretrained donors with no training data or gradient updates. source
  • ·Each claim in Agora is an immutable Git commit, and the paper says a controlled comparison is still needed. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire