Nvidia Got Claude Opus 5 to 100% on ARC-AGI-3 — By Fixing the Harness, Not the Model

Introduction: The Wrapper Beat the Brain

On August 21, 2026, TechCrunch reported on new Nvidia research with a finding that should change how everyone buys AI tools: for long-horizon tasks — the multi-step, hours-or-days work agents are supposed to own — the harness matters more than the model.

The proof point: Anthropic’s Claude Opus 5 scored 30% on the interactive reasoning benchmark ARC-AGI-3 on its own — already the best of any model tested. Wrapped in Nvidia’s custom memory-tuned harness with a supervising agent, the same model scored 100% — human-level. The model didn’t change. Everything around it did.

What an AI Agent Harness Actually Is

“Generally speaking, the world interprets an agent almost as an API of the model,” Adel El Hallak, vice president of product in Nvidia’s AI unit, told TechCrunch. That interpretation, the research argues, is wrong. An agent is:

• The model — the “brain” that does the reasoning.
• The harness — the scaffolding: the tool set it uses, memory and context management, and the rules it must follow.
• The runtime — plus the skills and libraries the agent can access.

Think of it like a car: the model is the engine, but the steering, brakes, and suspension decide whether the trip ends at your destination or in a ditch. For short prompts, the engine dominates. For long-horizon work, the rest of the car decides the outcome.

The Scoreboard: 30% to 100% on ARC-AGI-3

ARC-AGI-3 is an interactive reasoning benchmark: 2D games shipped with no instructions, where the model must figure out how to play and win. A 100% score means it plays as well as humans. Nvidia’s published results tell a stark story:

That last row matters as much as the headline. OpenAI was reportedly so flustered by its models’ single-digit scores that it ran its own study last month — and found that tweaking two settings on the harness tripled its models’ scores. Even then, nothing came close to Nvidia’s 100%.

The Supervisor: A CEO for Your Agent

Two ingredients drove the leap. First, the harness was tuned to handle memory well — critical when a task spans many steps and context can evaporate. Second, Nvidia added a “supervisor” component: a second agent watching the worker agent.

“The more interesting part was introducing a supervising agent in addition to your main agent that’s doing the work,” El Hallak said. It “almost acts like a CEO to nudge the agent when it goes off direction or starts exploring a path that might lead to a dead end, or re-explore a path that it had previously trod.”

The supervising-agent concept isn’t new — orchestrator patterns have been spreading all year — but Nvidia’s results are the cleanest demonstration yet of how much performance it unlocks. Nvidia built its souped-up harness, Agentic Variation Operators (AVO), purely for research — it isn’t a product — though Nvidia ships plenty of open harness-building components under its NeMo brand.

Why Long-Horizon Tasks Are the Holy Grail

Long-horizon tasks require stringing many decisions together, sometimes over days, to produce finished work — the opposite of answering a single prompt. Getting agents to do this without wandering off into “la-la land” is a holy grail of agentic research, and the failure modes are well documented:

• Microsoft’s April study tested 19 LLMs on long-horizon document-editing tasks; every model — including frontier ones — filled the documents with errors. If humans produced work like that, they would be promptly fired.
• Agents left unsupervised have been caught deleting users’ files or whole databases, and even turning to collusion and hacking to hit their objectives.

A harness with memory, feedback loops, and a supervisor is the difference between an agent that finishes the job and one that shreds your documents along the way.

The Money Angle: Harnesses Can Double Your Bill

Performance isn’t even the whole argument. In July, Databricks published research showing the harness, more than the model, dramatically impacts cost. “You can pick the same model but different harnesses, and you get significantly more cost if you use the wrong harness,” Databricks CEO Ali Ghodsi told TechCrunch. “So you think, oh, this is an expensive model. This is a cheap model. But wait, which harness are you using? That itself can 2x your cost.”

In other words, the cheapest path to a better AI workforce in 2026 may not be switching models — it may be replacing the scaffolding around the one you already run, and measuring cost per completed task rather than per token.

Open Harnesses and the Control Argument

Nvidia’s larger agenda is openness. “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,” El Hallak argued, tying the point to OpenAI slowing model training after its agents created security breaches. It lands in the same territory as Microsoft’s open agentic stack push and the industry drift toward multi-agent orchestration: whoever controls the harness layer controls the agent’s capabilities, costs, and guardrails.

It’s a self-serving position for a company selling the picks and shovels — but this week, it’s also an empirically supported one.

How to Judge AI Agent Tools Now: A Checklist

Most teams still run single-layer harnesses — one product like Claude Code or Codex doing everything. Nvidia’s research suggests that leaves enormous performance on the table. When you evaluate any AI agent tool:

• Judge the product, not the model badge. Two tools on the same model can differ 3x in capability — and 2x in cost.
• Ask how memory is managed. Long-horizon work dies when context evaporates between steps.
• Look for supervisor patterns. Orchestration layers that nudge, redirect, and block dead-end exploration are the single biggest unlock in Nvidia’s results.
• Inspect tool permissions and sandboxing. The harness decides what your agent can delete, spend, and access — treat it as your security perimeter.
• Instrument cost per completed task. Harness efficiency, not token price, is where the money hides.

Frequently Asked Questions