AI agents are showing real progress, but not all progress points in the same direction. NVIDIA’s AVO system reached a perfect score on a demanding benchmark and spent seven days improving computer code without stopping. AMD also reported a major rise in the number of software issues its agents fix automatically.
At the same time, a Princeton University study found that AI agents still struggle with the open-ended creativity and judgment needed for original research. Put together, these results point to a clear divide: agents can become powerful specialists before they become reliable independent researchers.
NVIDIA’s AVO pushes autonomous work further
NVIDIA AVO reached a 100.00 RHAE score on the ARC-AGI-3 benchmark, completing all 183 levels across 25 environments. It also used 12% fewer environment actions than VISTA, showing that the system reached the finish line with fewer steps.
AVO stands for Agentic Variation Operators. The architecture combines persistent memory, supervision, and tool use inside a general-purpose agent system designed for long-horizon tasks. In practical terms, that means it can keep working across a long sequence of actions instead of treating each task as a short, separate exchange.
The project also lifted Claude Opus 5 from a 30% model baseline to 100%. That result matters because the system did not rely only on a stronger starting model. Its architecture helped organize the work, maintain progress, and use tools across the benchmark’s environments.
AVO’s most demanding test involved GPU-kernel optimization. The system operated continuously for seven days, explored more than 500 optimization directions, and produced 40 committed kernel versions. Each version represented a path the system chose to keep as it searched for better performance.
On NVIDIA DGX B200 systems, the final multihead attention kernels outperformed cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% across the evaluated configurations. AVO then adapted the evolved kernel to grouped-query attention in approximately 30 minutes of additional autonomous work.
That sequence shows where autonomous agents already look useful: a defined goal, measurable results, access to tools, and enough time to test many alternatives. The system can explore, compare, keep successful versions, and adapt its work to a related problem.
AMD sees a similar rise in software engineering
AMD’s Radeon Software eXperience, or RSX, offers a separate example of agents becoming useful inside software development. AI agents initially fixed only 6% of RSX issues, but that figure rose to more than 75% by June 2026.
The percentage of issues fixed automatically reached 75 percent in June 2026. That change suggests agents can handle a growing share of engineering work when the problem sits inside an existing software process and the expected outcome is clear.
AMD also discussed future AI agent swarms. The idea involves multiple agents working together on engineering tasks, although the available results focus on the progress already seen in RSX rather than a completed swarm system.
The NVIDIA and AMD examples share an important feature. Neither requires an agent to invent an entirely new field of knowledge. The work involves benchmarks, code, experiments, optimization targets, and software issues that people can evaluate against clear standards.
Open-ended research remains a harder test
A six-day study by Princeton University researchers tested whether AI agents could move beyond engineering tasks and produce original AI research. The system received $3,000 in API credits and a GPU budget for experiments.
Anthropic’s Claude Opus 4.8 answered research questions drawn from two unpublished papers, with both papers submitted to NeurIPS 2026. The original authors rejected both papers. The agents reviewed literature, ran hundreds of experiments, and compiled the results, but the work fell far short of the quality expected at a top AI conference.
The agents struggled with the parts of research that are difficult to measure in advance. They had trouble exploring ideas, changing direction after weak results, and producing novel contributions. They also failed to use feedback from subagents or external AI reviewing tools, and they could not make effective use of available resources or follow detailed instructions.
The study found that the agents could solve engineering problems needed for AI research, but they lacked the judgment and creativity needed to create original research at the level of top conference papers. The agents did not engage in reward hacking, and the orchestrator agent caught hallucinations or misrepresentations from subagents.
That distinction helps explain why a system can optimize kernels for seven days yet still struggle to decide which research idea deserves seven days of work. Optimization starts with a target. Open-ended research must often discover the target, reject weak assumptions, and make a creative leap before any experiment can prove the idea useful.
The broader lesson is that AI progress may split into two paths. Narrow tasks can advance faster, especially when agents have tools, feedback, and clear measurements. Major advances in AI have also required creative leaps tied to transformers and new architectures, and that kind of originality may be necessary for recursive self-improvement.
For now, AVO and AMD’s RSX results show agents taking on longer and more valuable technical work. The Princeton study sets a boundary around that progress: powerful execution is not the same as independent discovery.
Based on
- NVIDIA AVO Reaches 100% on ARC-AGI-3, Demonstrating a Frontier-Level General-Purpose Architecture for Long-Horizon Autonomous Agents — developer.nvidia.com
- AME Agent Swarms Quietly Rewrite the Workflow – IEEE Spectrum — spectrum.ieee.org
- AI’s recursive self-improvement might not come so quickly after all | MIT Technology Review — technologyreview.com
