On August 21, 2026, Nvidia published a result that reframes how enterprise teams should evaluate AI agents: the same Claude Opus 5 model that scored 30% on the ARC-AGI-3 benchmark scored 100% when run inside Nvidia’s AVO agent system. The model did not change. The harness did.
For enterprise AI builders who have spent the past year debating which frontier model to standardize on, this is not a minor research footnote. It is a direct challenge to the assumption that model selection is the primary determinant of agent performance on complex, long-running tasks.
What happened
Nvidia’s Agentic Variation Operators (AVO) research team, led by principal engineer Terry Chen, published a technical blog post on the Nvidia Developer Blog describing how AVO achieved a 100.00 RHAE (Relative Human Action Efficiency) score on the ARC-AGI-3 public benchmark. The system completed all 183 levels across all 25 environments, using 6,624 environment actions. For reference, VISTA, a separate agent harness from MIT researchers that also reached 100%, required 7,542 actions on the same set. AVO used approximately 12% fewer.
ARC Prize, the nonprofit body that maintains the ARC-AGI-3 benchmark, separately reported Claude Opus 5’s baseline score at 30.2% when run at high reasoning effort without a custom harness. That 30% to 100% jump represents a threefold performance increase produced entirely by system design, not by switching to a more capable model.
The research team described the core finding directly: “Evaluating a model is not the same as evaluating an agent. Model capability matters enormously, but the surrounding system determines how effectively that capability can be converted into sustained autonomous progress.”
What ARC-AGI-3 actually tests
ARC-AGI-3 presents agents with unfamiliar, turn-based 2D game environments. No instructions are provided. No rules are stated. No objectives are named. An agent must explore through interaction, infer what its available actions do, discover what winning means, and then execute efficiently across progressively harder levels.
When the benchmark launched in March 2026, humans scored 100% and frontier AI systems scored 0.51%. By the end of August, a series of purpose-built harnesses (Tycho in late July, VISTA in early August, and now AVO) have each reached a perfect public-set score. The private-set benchmark, which withholds environments not available in the public set, remains unsolved.
The public-set result matters because it reveals something about architecture, not just about any single model. Three independent teams reached 100% by wrapping a frontier model in persistent state management, grounded feedback loops, and recovery mechanisms. None achieved it with a raw model call.
AVO’s four components
The Nvidia team identified four elements that determine long-horizon agent performance. These apply beyond ARC-AGI-3:
Persistent memory. The system carries forward prior implementations, test results, profiler outputs, and accumulated reasoning across the full task horizon. The agent resumes from its current state rather than reconstructing the problem from scratch on each invocation.
Tools. In GPU optimization, the tools were compilers, profilers, and test runners. In ARC-AGI-3, the tools were environment observation actions. The execution loop is the same: form a hypothesis, act, observe evidence, update state, and continue. Only the interface changes.
Grounded feedback. The system grounds decisions in concrete outcome signals, not in model-generated explanations of its own behavior. Compilers either pass or fail. Environments either respond or reject. This prevents the model from hallucinating progress it has not actually made.
Supervision. A second software module, the supervisor, monitors the main agent’s search trajectory. When the main agent drifts toward a dead end, revisits territory it already explored, or stops making progress, the supervisor intervenes and redirects it. This is the component the Nvidia team described as the key differentiator. “It almost acts like a CEO to nudge the agent when it goes off direction,” Nvidia engineer El Hallak told TechCrunch.
The GPU kernel result: proof of generality
The ARC-AGI-3 score is striking, but the more important proof of AVO’s architecture is the GPU kernel optimization result published alongside it. AVO ran autonomously for seven days on a GPU-kernel optimization task. It explored more than 500 optimization directions, committed 40 kernel versions, and produced multihead attention kernels that outperformed cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% on Nvidia DGX B200 systems. The system then adapted the evolved kernel to grouped-query attention in approximately 30 minutes of additional autonomous work.
The same architecture produced both results. The GPU optimization task involves compilers, profilers, and test runners. ARC-AGI-3 involves game-like environments with no documented rules. The feedback interfaces differ completely. The underlying agent loop is the same.
This is the result that matters most for enterprise buyers. A system designed around hypothesis formation, evidence accumulation, grounded feedback, and recovery does not need to be retrained or redesigned for each new task domain. The environment-specific tools change; the loop does not.
ARC-AGI-3 public set results at a glance
| System | Base Model | RHAE Score | Environment Actions |
|---|---|---|---|
| Tycho (July 2026) | Not disclosed | 100.00 | Not reported |
| VISTA (August 2026) | Claude Opus 5 | 100.00 | 7,542 |
| Nvidia AVO (August 2026) | Claude Opus 5 | 100.00 | 6,624 |
| Claude Opus 5 (no harness) | Claude Opus 5 | 30.2% | N/A |
Source: Nvidia Developer Blog, ARC Prize Foundation, TechCrunch.
Note: The ARC-AGI-3 semi-private and fully private competition sets have not been solved by any system. Public-set results do not carry over to those environments automatically.
What this means for enterprise AI builders
Most enterprise teams still evaluate AI vendors the same way they evaluate software vendors: benchmark the model, compare outputs, pick the highest score. This approach made sense when models were used for single-turn tasks. It breaks down for multi-step, long-running workflows, which is where most enterprise value actually sits.
The Nvidia result does not say that model selection is irrelevant. Claude Opus 5 at 30% was described as the best among models tested without a harness. A weaker model inside the same AVO system would produce a lower result. But the gap between the best model without a harness and the best model with one is approximately 70 percentage points on a demanding benchmark. That is the gap enterprise teams are leaving on the table when they optimize model selection and ignore runtime architecture.
Specifically, enterprise builders should evaluate the following:
Memory policies. How does the system preserve context across steps? Can it recover from a partial failure without restarting the full task? Does it compress context correctly when approaching token limits, or does it silently lose information?
Tool composition. Which tools does the agent have access to, and when is each available? Broad tool access increases capability. Unconstrained tool access increases risk. The permission model matters as much as the capability set.
Supervisor logic. Does the system have any mechanism to detect when the main agent is stuck or drifting? Without a supervisor layer, agents on long-horizon tasks tend to either repeat steps or pursue dead ends at full cost. A supervisor adds latency and token use but prevents unbounded waste.
Termination rules. How does the system decide when to stop, escalate to a human, or request clarification? An autonomous agent that cannot recognize the boundary of its own competence is a cost control problem.
Feedback grounding. Are the agent’s progress signals tied to actual external outcomes (test passes, API responses, database writes) or to its own generated reasoning? Grounded feedback prevents hallucinated progress and is the mechanism that allows AVO to run for seven days on a real engineering task without human intervention.
What AVO is not
AVO is a research architecture. Nvidia has not released it as a packaged product. The company’s commercially available agent building blocks sit in NeMo, with varying degrees of openness. Teams that want to apply AVO’s principles today will need to implement them using available frameworks rather than deploying AVO itself.
The result also covers only the public set of ARC-AGI-3. As RuntimeWire noted, the withheld private-set environments are the harder test of whether an architecture generalizes or has simply accumulated useful knowledge about publicly available games. No system has solved them yet.
The cross-model result using GPT-5.6 Sol is preliminary. Nvidia compared AVO-with-Opus and AVO-with-Sol on a subset of ARC-AGI-3 games and found complementary performance profiles, but the team did not publish full results and explicitly deferred a systematic comparison to future work.
What to watch
Enterprise teams tracking this space should watch for three developments. First, whether Nvidia publishes AVO’s implementation details in enough depth for independent reproduction and third-party benchmarking. Results from a single lab do not establish architecture generality on their own.
Second, whether the supervisor layer pattern (a monitoring agent that redirects the primary agent) appears in general-purpose commercial products. The concept is not new in theory. It is new in demonstrating material benchmark impact at this scale.
Third, whether the ARC-AGI-3 private-set problem is solved in the next quarter. A system that transfers from GPU optimization to interactive reasoning is interesting. A system that transfers to novel, unseen environments would be significant. Those environments are the ones that most resemble real enterprise deployments, where the playbook is never written out in advance.
The harness is now part of the product. Teams that have not yet designed their agent runtime with the same care they apply to model selection are operating with a structural disadvantage that this result makes visible.
Enera works with enterprise teams to design and deploy production-ready AI agent systems. If your team is moving from model evaluation to system architecture, reach out here. For more on the enterprise AI infrastructure that enables long-horizon agents, see our coverage of Anthropic’s Computer Use and Skills API going GA and Claude Opus 5’s enterprise pricing.