AVO cleared the benchmark. NVIDIA’s system reached a 100.00 RHAE score on ARC-AGI-3, completing all 183 levels across 25 environments with 12% fewer environment actions than VISTA. That result positions AVO as a system-level general-purpose architecture for long-horizon autonomous agents—not another model waiting for a carefully scripted task.
The research project elevates Claude Opus 5 from a 30% model baseline to 100%. The distinction matters because AVO’s result comes from the full system, including its ability to plan, test, adapt, and continue working across many iterations without requiring each step to be manually prescribed.
AVO Turns Model Output Into an Engineering Loop
AVO operated for seven days, explored more than 500 optimization directions, and produced 40 committed kernel versions. That is a long engineering cycle by any standard, and it shows the system sustaining a productive loop instead of producing one impressive answer before wandering off into the digital shrubbery.
The work focused on multihead attention kernels running on NVIDIA DGX B200 systems. The resulting kernels outperformed cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5%, giving the autonomous process a measurable result beyond benchmark completion.
AVO also adapted the evolved kernel to grouped-query attention in approximately 30 minutes of additional autonomous work. That step matters because it shows the system applying its work to another attention configuration without requiring every move to come from a human operator.
The combination of benchmark coverage, lower environment-action use, extended operation, and kernel performance creates a more demanding picture of agent capability. AVO did not only complete 183 levels; it also spent seven days exploring alternatives and delivered code versions that improved performance against established comparisons.
AMD Reports Agents Moving Into Production Software
AMD’s software results point in the same direction from a different angle. In AMD’s Radeon Software eXperience, the percentage of software issues fixed automatically by AI agents reached 75 percent in June 2026, up from 6 percent in October 2025 when out-of-the-box AI tools resolved only 6% of issues.
AMD has also surpassed a 25% productivity boost from AI in software development, reaching a 30% overall boost. In some software components, more than 80% of the code is now generated using AI, while AI agent swarms can develop solutions independently.
The timeline is the important part. AMD’s figures move from limited results in October 2025 to 75% of RSX issues fixed automatically in June 2026, while the broader development effort reaches a 30% productivity boost. That is not a promise about a distant workplace; it is a measured shift in how software work gets completed.
Together, NVIDIA AVO and AMD’s software results describe AI agents as systems that can carry work across multiple stages: exploring options, committing versions, adapting code, fixing issues, and developing solutions independently. The model remains part of the machinery, but the useful unit is the complete operating loop.
That is also why AVO’s 100% score needs careful reading. The result does not turn Claude Opus 5 into a 100% model baseline; it elevates a 30% model baseline to 100% through the surrounding architecture. The impressive part is not a single model suddenly becoming omniscient. It is the system refusing to stop at the first plausible answer.
As of Aug 21, 2026, the evidence points to a transition from AI assistance toward autonomous engineering systems. AVO’s seven-day run and AMD’s 75% issue-resolution figure show the same pressure from opposite ends: agents are being judged less by what they can generate once and more by what they can finish over time.
Based on
