NVIDIA has developed the Agentic Variation Operators (AVO) architecture, a general-purpose agent system designed for sustained autonomous operation on long-horizon tasks. This architecture integrates persistent memory, supervision, and tool-use. The AVO system was applied to the ARC-AGI-3 benchmark, where it achieved a 100.00 RHAE score.
This performance involved completing all 183 levels across 25 environments. The AVO system demonstrated this capability with 12% fewer environment actions than VISTA, another agent system. The research indicates that system design, rather than model capability alone, can enable frontier-level long-horizon performance.
Key Points
- The NVIDIA AVO architecture is a general-purpose agent system for long-horizon tasks.
- AVO integrates persistent memory, supervision, and tool-use.
- The system achieved a 100.00 RHAE score on the ARC-AGI-3 benchmark.
- AVO completed all 183 levels across 25 environments in the ARC-AGI-3 public set.
- It used 12% fewer environment actions than VISTA on the ARC-AGI-3 benchmark.
- The research elevated Claude Opus 5 from a 30% model baseline to 100% within the complete AVO system.
- In GPU-kernel optimization, AVO operated continuously for seven days, explored over 500 optimization directions, and produced 40 committed kernel versions.
- On NVIDIA DGX B200 systems, AVO's optimized multihead attention kernels outperformed cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5%.
Context
According to the NVIDIA Developer Blog, a frontier language model is only one component of an AI agent. The surrounding agent system, or harness, determines how the model receives context, uses tools, maintains state, responds to feedback, recovers from failure, and sustains progress over long-running tasks. The AVO architecture addresses the challenge of building a general-purpose agent architecture that enables frontier language models to work reliably on extended, multistep tasks.
AVO's design focuses on sustained autonomous operation across long horizons. It replaces the predefined variation step of conventional evolutionary-search systems with an autonomous agent that decides what to inspect, change, test, and commit. For the ARC-AGI-3 benchmark, the same general-purpose agent was connected to a different task interface, with only environment-specific tools and evaluation changing. The system's ability to preserve progress beyond a single model context relies on persistent memory and supervision. Persistent memory carries forward prior implementations, evaluation results, and reasoning, while a supervisor monitors the broader trajectory for stagnation and can redirect the main agent.
Why It Matters
This development indicates that system-level architecture plays a significant role in achieving high performance and generality in agent systems, beyond the capabilities of the underlying language model. Builders and researchers can note that effective agent design, including mechanisms for persistent memory and supervision, can significantly enhance the performance of existing models on complex, long-horizon tasks.
What To Do
- Note the emphasis on system design over model capability alone for long-horizon tasks.
- Consider how persistent memory and supervisory mechanisms could be integrated into agent architectures.
- Review the described application of AVO to GPU-kernel optimization as an example of sustained autonomous work.
- Examine the ARC-AGI-3 benchmark results to understand the performance implications of the AVO architecture.
Keep Exploring
/atlas/claude-family