Nvidia's AVO Harness Takes Claude Opus 5 from 30% to 100% on ARC-AGI-3

Nvidia published research on August 21 showing that the software harness surrounding an AI model matters more than the model itself when agents must execute long-horizon tasks. The proof: wrapping Claude Opus 5 in a custom harness called Agentic Variation Operators (AVO), Nvidia researchers pushed its ARC-AGI-3 score from 30% — already the highest raw score among all models tested — to 100%. ARC-AGI-3 consists of 2D games with no instructions, where the agent must infer the rules and win, the same way a human would.

This result extends a pattern that has been building across pipeline research in 2026. In July, Databricks CEO Ali Ghodsi reported that choosing the wrong harness for the same model can 2x inference costs — a multiplier that has nothing to do with model selection. Together, the two data points reframe harness architecture as the dominant variable in agentic system performance, not parameter count or base benchmark capability.

What AVO Does Differently

AVO introduces two structural changes most production agent deployments lack. First, it manages memory and context explicitly, preventing the drift that causes agents to abandon coherent reasoning mid-task. Second, it adds a supervisor component — a separate agent layer that monitors the primary working agent and intervenes when it stalls, loops, or pursues dead ends. Nvidia VP of Product Adel El Hallak described the supervisor as functioning "almost like a CEO," redirecting the agent when it goes off course or revisits already-explored paths.

The two-layer architecture is not conceptually novel, but El Hallak noted that most practitioners currently rely on a single harness layer — tools like Claude Code, Codex, or Hermes — without any supervisory component above it. The gap between that common configuration and what AVO implements is precisely where the 30%-to-100% jump originates. The control layer question is one the field has circled without resolving into standard practice.

Benchmark Context

ARC-AGI-3 has been a pointed benchmark for frontier labs. OpenAI's models scored below 10% on it, prompting the lab to run its own harness investigation last month. OpenAI found that tweaking just two harness settings tripled its models' scores — but none came close to 100%, which underscores that harness ceiling gains are not yet exhausted.

Microsoft's April research adds further context: testing 19 LLMs on long-horizon document editing tasks, Microsoft found that every model — including frontier ones — introduced errors at a rate unacceptable from a human worker. Harness-level memory management and supervisory redirection are engineered to address exactly that failure mode.

Openness as the Strategic Argument

AVO is not a commercial product. Nvidia ships harness tooling under the NeMo brand, with a mix of commercial and openly available components. El Hallak tied the open-harness argument directly to security: "We believe in having an open agent stack — where you have control across the harness, across the infrastructure, across the runtime — is what's required for us to usher the ecosystem forward and securely." He pointed to OpenAI slowing model training due to models creating security breaches as an illustration of why full-stack control matters. The security surface that agentic systems expose grows in proportion to how much of the stack is opaque to the operator.

For practitioners, the finding reframes a core allocation question. If the harness contributes more to task-completion accuracy and cost than the base model, and if harness components are increasingly open and composable, engineering investment in scaffolding — memory architecture, supervisor layers, tool orchestration — carries higher expected return than model selection alone. Nvidia's research provides a concrete, quantified data point that is difficult to dismiss: 30% to 100% on a benchmark frontier models collectively struggle with, achieved without changing the model at all.