NVIDIA has revealed AVO, an agent system built to handle long and complex programming jobs. If you build with AI, this shows that tools and feedback loops matter just as much as raw model power.
How the system works
AVO stands for Agentic Variation Operators. Instead of giving one answer and quitting, the system works through a continuous loop. It inspects tasks, tests code, checks feedback from the environment, and uses persistent memory to correct mistakes. The system does not stop.
NVIDIA shared two big test results on the NVIDIA Technical Blog:
- GPU kernel optimization. AVO explored more than 500 directions and committed 40 kernel versions. It ran up to 10.5% faster than FlashAttention-4 on NVIDIA DGX B200 systems.
- ARC-AGI-3 benchmark. On the public task set, AVO scored a 100.00 RHAE rating. It cleared all 183 levels across 25 environments while using 12% fewer actions than VISTA. By comparison, a standalone Claude Opus 5 baseline scored only 30%.
The numbers you shouldn’t trust yet
All these figures come directly from NVIDIA’s own technical post. No outside researchers have verified them yet. NVIDIA also updated its post to note that the 100% score applies only to the public ARC-AGI-3 set, not the semi-private or private competition tasks. These tests do not prove that AVO is cheaper, safer, or ready for your live codebase. Extra supervisory tools also add compute costs and potential failure points.
Why this shifts your AI choice
Until now, teams mostly picked AI tools by looking at static model leaderboards. AVO argues that real performance lives in the harness around the model. Good scaffolding makes a huge difference. The real question is whether an agent can use tools, recover from bugs, and reach a solid result without wasting compute.
What to test today
Do not rush into production. If you want to evaluate an agent system, start with a disposable coding task or the public ARC-AGI-3 set. Lock down your model, tools, memory format, action budget, and hardware. Then run a base model next to the full agent loop. Measure runtime, token use, retries, and total costs before you commit your engineering pipeline.