Nvidia published research on Friday arguing that the scaffolding wrapped around a language model matters more than the model itself once an agent has to work for hours instead of seconds. The claim arrives with an eye-catching number attached: a custom harness the company calls Agentic Variation Operators, or AVO, carried Claude Opus 5 to a perfect score on ARC-AGI-3, a benchmark the same model clears only 30% of the time on its own.
ARC-AGI-3 is deliberately unfriendly to pattern matching. It hands an agent a set of 2D games with no rules, no tutorial and no goal description, then asks it to work out what winning looks like the way a curious human would. Frontier systems have struggled badly with it — OpenAI's entries landed under 10% — which is part of why the result drew attention within hours of being posted.
What the harness actually adds
A harness is everything that surrounds the raw weights: tool definitions, the memory layer, the runtime, and the rules that decide what the agent sees next. Adel El Hallak, who runs product for Nvidia's AI unit, told TechCrunch that the industry tends to collapse all of that into the model itself, treating an agent as little more than an API call. In his framing, an agent is the model plus the scaffolding plus the runtime plus whatever skills and libraries it can reach.
The piece Nvidia credits for most of the gain is a second agent sitting above the first. El Hallak described it as a CEO-like supervisor that steps in when the working agent stalls, wanders toward a dead end, or starts re-walking ground it has already covered. Most production setups today — Claude Code, Codex, Hermes — run a single layer with no such oversight.
Read the fine print on 100%
The perfect score applies to ARC-AGI-3's 25-game public set, not the semi-private or private evaluations that decide the official leaderboard, where the verified high currently sits at 40%. Overfitting to a public set is a real hazard, and AVO has not been submitted for competition scoring. Nvidia also reports that the harness needed 12% fewer environment actions than bare Opus 5, and that pairing AVO with GPT-5.6 Sol on a harder subset produced a different profile — quicker wall-clock progress, more actions consumed.
AVO did not begin life as a benchmark runner. The accompanying paper, credited to 23 researchers, describes it as an evolutionary search method that swaps hand-designed mutation and crossover operators for coding agents able to inspect a candidate's lineage, consult a domain knowledge base and verify their own edits. Pointed at attention kernels on Blackwell B200 hardware for seven days of unattended evolution, it produced implementations up to 3.5% faster than cuDNN and up to 10.5% faster than FlashAttention-4.
A pattern, not a one-off
Nvidia's finding lines up with a run of similar results across the industry. OpenAI, stung by its ARC-AGI-3 showing, reported last month that adjusting two harness settings tripled its scores. Microsoft tested 19 models on long-horizon document editing in April and found every one of them, frontier systems included, riddling the output with errors. Databricks measured the cost dimension in July, with chief executive Ali Ghodsi noting that the same model on the wrong harness can double what a workload bills.
AVO is not a product. Nvidia sells and open-sources harness components under the Nemo brand, and the company's larger argument is that open scaffolding hands builders far more levers than most of them realise they have. Whether the perfect score survives contact with the private evaluation set is the next thing worth watching.
