NVIDIA Open-Sources Agora: Git as Shared Memory for 13 Autonomous Research Agents

NVIDIA's team (Yifan Zhang, Jan Kautz, Yi Dong, et al.) released and open-sourced Agora (arXiv:2609.18094), positioned as "Git as Shared Memory for Collective AutoResearch." It addresses the problem of multiple coding agents duplicating each other's experiments when researching the same question by turning Git into shared memory: the entire research process is recorded as an append-only DAG in Git, where every result, hypothesis, and verification is an immutable commit.

Confirmed

• No task assignment, no central planner; agents collaborate autonomously
• 13 agents ran for nearly 12 days, producing 1,703 contributions
• 165 independently reproduced results (per @markk, @dairai, @alexverem)
• Paper public, system open-sourced (announced via @nvidia's official account)

Why it matters

Agora demonstrates a decentralized multi-agent collaboration paradigm: shared knowledge through immutable version records rather than a scheduler, letting autonomous research agents accumulate, verify, and reuse each other's results—reducing duplicated work and offering a reusable engineering blueprint for large-scale automated scientific research.

2026-09-17 ~ 2026-09-19 · 6 related posts

• [source] Nvidia's Agora: 13 LLM Agents Run 12 Days Unsupervised via Git-Based Shared Memory — nvidia · 2026-09-17
• NVIDIA Agora: 13 autonomous agents used Git as shared memory for 12 days, making 1,703 contributions — mark_k · 2026-09-19
• Agora Paper: Git as Shared Memory Lets 13 Agents Close 62% of Gap to Trained GPT-2 — alex_verem · 2026-09-19

3 near-duplicate retellings: dair_ai · _akhaliq · alex_verem