Nvidia AVO hit 100% on ARC-AGI-3. The harness did it.

The same model, two scores: about 30% and 100%. Nvidia published results on August 21 showing its research agent architecture, AVO, clearing every level of the ARC-AGI-3 public set — running on Claude Opus 5, the model that scores roughly 30% on that benchmark by itself. Nothing about the model changed. The agent harness around it did all the work.

What actually happened

Per Nvidia's technical blog, AVO — Agentic Variation Operators — posted a 100.00 RHAE score across all 25 environments in the ARC-AGI-3 public set, completing all 183 levels. Claude Opus 5 at high reasoning effort lands near 30% on the same set unassisted. AVO used 6,624 environment actions against 7,542 for VISTA, the prior top result, roughly 12% fewer.

Three pieces do the lifting. Persistent memory that carries prior attempts, evaluation results, and profiler output across iterations. An inspect-plan-implement-evaluate loop that runs until something passes. And a supervisor watching the whole trajectory for stagnation and repeated unproductive cycles, so the agent gets pulled out of a loop instead of burning tokens inside it.

The same architecture spent seven days on GPU kernel optimization, explored more than 500 directions, and committed 40 kernel versions — beating cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% on the configurations tested.

Two caveats Nvidia states plainly and the headlines drop: this is the public set, not the semi-private or private competition sets, and the VISTA comparison is not a controlled ablation. There is no open-source release. AVO is a research result, not something you can install.

Why the agent harness matters for your business

Stop shopping for models. Start building scaffolding. A 30-to-100 delta from architecture alone is the clearest statement yet that most teams are tuning the wrong layer. If your agent is failing on multi-step work, the next model release probably will not fix it.

Persistent memory is not a vector database. What AVO carries forward is structured: what was tried, what the evaluator said, what the tooling printed. That is a run log with a schema, not embeddings of your docs. It is also the cheapest thing on this list to build.

The supervisor is the part nobody builds and everybody needs. An agent that retries the same failing approach eleven times costs real money and produces nothing. A loop counter, a no-progress detector, and a hard stop are an afternoon of work. We ship them by default now because we have paid the other bill.

Write the evaluator first. AVO works because it can tell pass from fail without a human. If your task has no automatic check — no test, no diff, no validator — you do not have an agent task yet. You have a draft generator with a person on the end.

Sources: NVIDIA Technical Blog, TechCrunch.