Nvidia's Agora: 13 LLM Agents Run 12 Days Unsupervised via Git-Based Shared Memory

nvidia · hf · 2026-09-17

Nvidia introduces Agora, a shared-memory system for autonomous research agents, built to fix the duplication problem when multiple agents search in parallel from scratch.

Design

Research is recorded as an append-only DAG in Git: every result, insight, hypothesis, verification, and report is an immutable commit, with parent edges showing dependencies; anyone can check out and rerun any claim
A derived index exposes the frontier, neglected branches, and verification status of each claim
A diversity-aware selection rule prevents collapse onto a single leading solution

First sustained run

13 LLM workers, with no assigned tasks and no central planner, ran nearly 12 days on a weight-transfer problem: initialize a frozen 119.6M-parameter attention-SSM hybrid (matching no donor) using 141 pretrained donor models, with no training data or gradient updates
1,703 contributions were published; the evaluator improved from 3.39 to 1.899 bits per byte, closing 62% of the gap to a trained GPT-2 124M
The winning recipe (compressing donor next-token statistics into embeddings and output head, plus sparse edits to inject short-range context) has a 145-commit ancestry across 15 accounts; 165 independent reproductions were posted, none failed

The authors also detail the single mid-run human intervention that broke a monoculture, and what the trace does and does not establish.

Original post →

Nvidia's Agora: 13 LLM Agents Run 12 Days Unsupervised via Git-Based Shared Memory

More from coding & agent