NVIDIA's Agora: 13 Git-Linked AI Agents Self-Organize 12 Days of Research, Closing 62% of Gap to GPT-2
alex_verem · x · 2026-09-19
NVIDIA researchers published Agora, a system that uses Git as shared memory for autonomous research agents: every result, hypothesis, and verification is an immutable commit in an append-only DAG, with a derived index exposing the frontier, neglected branches, and verification status.
Key facts from the first sustained run:
- 13 agents (Claude Code on Opus 4.7, Codex on GPT-5.5), no central planner, no assigned tasks, running 12 days
- Task: initialize a 119.6M-parameter attention-SSM hybrid from 141 pretrained donor models with mismatched dimensions, no training data or gradient updates allowed
- 1,703 contributions published: 1,124 experiments, 284 observations, 203 hypotheses, 165 cross-checks with zero failures
- Evaluator score improved from 3.39 to 1.899 bits/byte, closing 62% of the gap to a properly trained GPT-2 124M
- The winning 145-step recipe was built mostly on other agents' work — no single agent produced it
Most striking: agents stalled for 5 days refining one idea (over a third of all activity in one cluster); after researchers introduced a map of activity coverage, one agent pivoted to a neglected direction and set a new best score within a day. Conclusion: coordination structure, not individual capability, was the bottleneck. The authors note they ran no control without the shared system, so superiority over solo agents remains unproven.
More from coding & agent
- 'I asked for 3 lines, got 400 lines and 17 tests': dev's complaint about AI coding agents goes viral — burny_tech · 2026-09-19
- Pace Layers explain AI anxiety: when slow layers are forced to sprint — dbreunig · 2026-09-19
- AWS treats DRAM as the scarce resource in agent infra with new AgentCore runtime — bookwormengr · 2026-09-19
- pdfcn: open-source React PDF components built on Takumi and Forme — tom_doerr · 2026-09-19
- Elvis Saravia on integrating Jev into agent harnesses: cheaper workflows, verifiers and new UX — omarsar0 · 2026-09-19
- Developer: AI short-circuits the doubt phase, from idea straight to prompting — dfinke · 2026-09-19