Agora ran 13 LLM agents for 12 days using a Git DAG as shared memory and closed 62% of the gap to GPT-2 124M

arXiv·high signalXBlueskyLinkedInCopy link

arXiv 2609.18094, submitted 2026-09-16, stores autonomous research as an append-only DAG of Git commits so agents can check out any prior commit to verify a claim or build on it, with a diversity index that stops the population converging on one approach. A 12-day run with 13 language-model agents on weight initialization produced 1,703 contributions and moved performance from 3.39 to 1.899 bits per byte, closing 62% of the gap to a trained GPT-2 124M. The winning solution traced back through 145 commits across 15 accounts, and 165 independent reproductions ran without failure.