AVO is Nvidia’s agent architecture for long-horizon autonomy, presented as a research project rather than a product. Its first reported week went to evolving GPU kernels on DGX B200 systems, exploring more than 500 directions to beat cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5%; transferred unchanged to ARC-AGI-3 with Claude Opus 5 inside, it swept all 183 levels from a 30% model baseline. In this corpus it is the clearest case of the harness generalizing while the weights stay fixed.
AISystem
NVIDIA AVO
Nvidia's general-purpose agent architecture for long-horizon autonomy, which lifted Claude Opus 5 to a perfect score on ARC-AGI-3 and evolved GPU kernels past FlashAttention-4.
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to August 23, 2026 (theinnermostloop.substack.com)
Development timeline
Newest reporting dates first. Reporting dates may differ from event dates.
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.