NVIDIA presents Agentic Variation Operators (AVO), a general-purpose agent architecture designed to sustain long-horizon autonomous work through persistent memory, tool use, and supervision. AVO was first exercised on demanding GPU-kernel optimization tasks and later applied to the interactive ARC‑AGI‑3 benchmark, where it achieved a 100.00 RHAE score on the public set.
What is AVO?
AVO (Agentic Variation Operators) is an NVIDIA-developed agent system for coding and long-horizon autonomous workflows. The agent can inspect and edit code, run commands, consult documentation, and validate results by execution. Its distinguishing aim is to preserve and compound progress across many iterations rather than produce single-response outputs.
In GPU-kernel optimization, AVO replaces the fixed variation step of traditional evolutionary search with an autonomous decision-maker that chooses what to inspect, change, test, and commit. For ARC‑AGI‑3, the same core agent was attached to a different task interface: the agent remained the same while environment-specific tools and evaluation changed.
GPU-kernel optimization: an early stress test
GPU-kernel optimization provided a challenging initial application because the search space is large and performance effects are unpredictable without execution. Key results reported by NVIDIA:
- AVO ran continuously for seven days on an attention-kernel study.
- It explored more than 500 optimization directions and produced 40 committed kernel versions.
- On NVIDIA DGX B200 systems, the evolved multihead attention kernels outperformed cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% in the evaluated configurations.
Beyond final kernel gains, the experiment demonstrates that AVO can sustain a productive engineering loop over many iterations without requiring each step to be manually scripted. NVIDIA emphasizes that building a trusted agent stack requires designing for performance, reliability, and security at the system level, not treating those as model-only properties.
Mechanisms for long-horizon agentic work
Two mechanisms are highlighted as critical for preserving progress across model invocations: persistent memory and supervision.
- Persistent memory carries forward prior implementations, evaluation results, compiler and profiler outputs, and accumulated reasoning so the agent can resume from current state rather than rebuilding the search repeatedly.
- The supervisor monitors the broader search trajectory for stagnation or unproductive cycles and can redirect the main agent to alternative strategies. During the seven-day run, the main agent decided what to inspect, change, test, and evaluate, while the supervisor helped maintain forward progress when the search plateaued.
These mechanisms enable the system to sustain work beyond a single model context.
Applying AVO to ARC‑AGI‑3
NVIDIA applied the same AVO architecture to ARC‑AGI‑3, an interactive reasoning benchmark in which agents enter unfamiliar game-like environments without instructions, explicit rules, or stated goals. The agent must explore by interaction, infer dynamics and objectives, and act efficiently across progressively harder levels.
ARC‑AGI‑3 uses Relative Human Action Efficiency (RHAE), combining task completion with per-level action efficiency relative to first-time human baselines. This makes ARC‑AGI‑3 a demanding long-horizon agent task: success requires preserving useful knowledge, learning from prior interactions, recovering from mistakes, and spending environment actions efficiently.
Design choices in the NVIDIA AVO ARC‑AGI‑3 setup included a direct-interaction approach (informed by VISTA principles) implemented independently rather than centering on explicit programmatic world-model construction as in Tycho. In AVO’s configuration the LLM received text-only observations: each observation was supplied as an exact 64 x 64 text grid (no images or image tokens). The agent also received the available actions without provided descriptions of game rules or goals and had to infer effects via interaction.
ARC‑AGI‑3 performance
AVO’s public-set results on ARC‑AGI‑3 (using Claude Opus 5):
- AVO achieved a 100.00 RHAE score across all 25 public environments, completing all 183 levels.
- The run used 6,624 environment actions in total.
- For comparison, VISTA reports 7,542 environment actions with Claude Opus 5 for the same 183 public-set levels; AVO used about 12% fewer actions in that cross-system comparison.
NVIDIA stresses these are not controlled ablations: the two systems differ in backend agent design, observation representation, memory, context management, and other implementation details. The AVO memory system, which carries forward useful understanding and reduces repeated exploration, may matter over long horizons, but this experiment does not isolate its individual effect.
The team also ran limited experiments pairing AVO with GPT‑5.6 Sol on a challenging subset of games. In those preliminary comparisons, Sol matched levels faster in wall-clock time in several cases, while Opus used fewer environment actions in matched-level comparisons, suggesting complementary operating profiles across models.
All reported results cover the 25-environment ARC‑AGI‑3 public set using the official scorecard and RHAE metric; they do not include semi-private or private competition sets.
Lessons learned
The central finding is that the same agent architecture transferred from specialized GPU-kernel optimization to very different interactive reasoning tasks. Although feedback channels and interfaces differ (compilers/profilers vs. environment transitions), the core agent loop is consistent: form hypotheses from incomplete evidence, act through an external interface, observe consequences, preserve useful state, revise the problem model, recover from incorrect assumptions, and continue across a long horizon.
This suggests that generality can arise not only from domain knowledge but from the machinery that allows reasoning and feedback to compound over time. Long-horizon capability should be considered a property of the entire system: memory, tools, feedback, and recovery shape what survives between steps and how progress is sustained.
Looking ahead
AVO research started with autonomous software engineering and high-performance GPU-kernel optimization; ARC‑AGI‑3 shows the architecture can transfer to very different reasoning environments. The broader opportunity is to build general-purpose agent systems around persistent state, tool use, grounded feedback, recovery, and long-horizon context management so they can accumulate evidence and sustain progress across diverse tasks. The model matters, but it is not the whole agent.
Related resources
The article references the AVO paper (AVO: Agentic Variation Operators for Autonomous Evolutionary Search), the ARC‑AGI‑3 benchmark and scoring methodology, and related projects VISTA and Tycho for further reading.
