Nvidia's Harness Lifts Claude Opus 5 From 30% to 100% on ARC-AGI-3
News Summary
On August 21, 2026, Nvidia published research arguing that the biggest gains in AI agent performance are now coming from the software layer wrapped around a model, not from the model itself. The company calls this wrapper a "harness" — the orchestration code, memory system, tool access, and supervisory logic that turns a raw language model into an agent capable of operating on its own for extended periods.
The core claim: the harness, not the model, drives performance
Adel El Hallak, Nvidia's VP of Product for agentic AI, said the industry has been thinking about agents the wrong way, treating an agent "almost as an API of the model" rather than as a system built from several interacting parts. In Nvidia's framing, an agent is the model plus the scaffolding around it — the harness — which includes the tools the model can call, the way memory is stored and retrieved, the rules that constrain its behavior, and the checks that keep it from going off track during long, multi-step tasks.
To make the point concrete, Nvidia researchers took Anthropic's Claude Opus 5 and ran it on ARC-AGI-3, a benchmark from the nonprofit ARC Prize Foundation that drops an AI agent into unfamiliar, game-like environments with no instructions or stated goals. The agent has to explore, form a working model of the rules, and complete objectives purely through trial and observation across many turns, which makes it a stringent test of planning and memory rather than one-shot question answering.
From 30% to 100%: the AVO experiment
Run with only its default settings, Claude Opus 5 completed roughly 30% of ARC-AGI-3's public levels. Nvidia then wrapped the same model in its own harness framework, called Agentic Variation Operators (AVO), adding persistent memory across turns and a "supervisor" component that Nvidia describes as functioning like a CEO — reviewing the agent's progress, spotting dead ends, and redirecting its strategy before it wastes further actions. With AVO in place and no other change to the underlying model, the system completed all 183 levels across the benchmark's 25 public environments, reaching a perfect 100.00 score on ARC-AGI-3's Relative Human Action Efficiency (RHAE) metric, and did so using about 12% fewer environment actions than the previous best-performing system on that leaderboard, a harness called VISTA.
Nvidia has been developing AVO since introducing it as a general-purpose coding agent system in late March 2026, originally built to optimize CUDA GPU kernels. It is worth noting that this ARC-AGI-3 result comes from Nvidia's own reimplementation of the benchmark's public task interface, run and reported by Nvidia's team, rather than from a run independently verified on the ARC Prize Foundation's official leaderboard, and it covers only the 25 publicly released environments rather than the benchmark's semi-private or fully private competition sets used for official rankings.
Corroborating evidence from across the industry
Nvidia's demonstration lands alongside a string of similar findings from other organizations over the past several months, according to the reporting. In April 2026, Microsoft tested 19 large language models on real-world document editing tasks and found that, without a well-designed harness around them, every model produced work riddled with errors. In July 2026, OpenAI reported that it was able to roughly triple its own models' scores on ARC-AGI-3 simply by adjusting two settings within the harness surrounding the model, without changing the model itself. That same month, Databricks published research finding that the choice of harness affected the total cost of running an agentic task about twice as much as the choice of underlying model did.
Nvidia's own technical blog has separately detailed a related framework, referred to internally as NOOA, that lays out six harness design capabilities engineers can use to boost model performance: typed inputs and outputs that get validated rather than passed as free text; letting models operate on live objects by reference instead of serialized data; treating written code as the action itself, complete with control flow; making the orchestration loop plain, editable code rather than a hidden black box; keeping durable, typed state on agent objects instead of burying it in conversation history; and exposing context and event history as APIs the model itself can inspect and manage. Nvidia has reported that this approach reached 82.2% accuracy on the SWE-bench Verified software-engineering benchmark using notably fewer model calls and tokens per task than comparable systems.
Why it matters
The findings reframe how the AI industry should think about progress in autonomous agents. Rather than treating every capability jump as a function of a newer, larger, or more expensive underlying model, Nvidia's research suggests that the engineering layer sitting on top of a model — how it remembers, how it is supervised, and how it is allowed to act — can be the larger lever for real-world task completion, particularly on long-horizon problems that unfold over many steps. For developers and enterprises building AI agents, that implies that investment in harness design, memory architecture, and supervisory logic may pay off as much as, or more than, upgrading to the newest model release.