What happens when an agent system turns a 30% model into a 100% score on ARC-AGI-3?

The agent that stayed on task for a week

A coding agent usually writes a snippet, runs a test, and stops. AVO kept going for seven days. It explored more than 500 optimization directions for an attention kernel. Forty versions were committed. Each one survived compilation, profiling, and correctness checks without a human stepping in.

The result on NVIDIA DGX B200 systems: up to 3.5% faster than cuDNN and up to 10.5% faster than FlashAttention-4 across the tested configurations. When the team asked the same agent to adapt the evolved kernel to grouped-query attention, it finished in roughly 30 minutes.

Memory that survives context limits

Most agents lose their thread when the model context fills up. AVO solves this with persistent memory that carries forward prior implementations, compiler output, profiler traces, and the reasoning that led to each decision. The main agent resumes from the current state instead of reconstructing the search from scratch.

A supervisor watches the broader trajectory. When progress plateaus or the agent repeats unproductive cycles, the supervisor redirects it toward alternative strategies. During the seven-day kernel run, the main agent still decided what to inspect, change, test, and evaluate. The supervisor only intervened when forward motion stalled.

Same loop, different world

GPU-kernel optimization feeds back through compilers, tests, and profilers. ARC-AGI-3 feeds back through environment transitions in unfamiliar game-like worlds. The interfaces differ. The loop does not.

In both settings the agent must form hypotheses from incomplete evidence, act through an external interface, observe consequences, preserve useful state, revise its model of the problem, recover from incorrect assumptions, and continue making progress over a long horizon. What transfers is not domain knowledge. It is the machinery for sustained autonomous progress.

Text grids and action counts

For ARC-AGI-3, the observation was an exact 64 by 64 text grid. No images. No image tokens. The agent received available actions without descriptions of rules or goals and had to infer effects through interaction.

Using Claude Opus 5, AVO completed all 183 levels across the 25 public-set environments with a 100.00 RHAE score. It took 6,624 environment actions. VISTA, a different harness also using Claude Opus 5, reported 7,542 actions for the same levels. AVO used approximately 12% fewer actions in this cross-system comparison.

The two systems differ in agent backend, observation representation, memory, context management, and other implementation details. The comparison is not a controlled ablation. One architectural difference that may matter over long horizons is the AVO memory system, designed to carry useful understanding forward and reduce repeated exploration. This experiment does not isolate its individual contribution.

Model versus system

ARC Prize separately reports roughly 30% for Claude Opus 5 at high reasoning effort. The AVO run used the same model family under a different reasoning setting and a substantially different agent system. These numbers should not be read as a direct measurement of AVO's contribution. They illustrate that model-level evaluation alone does not characterize the performance of a complete agent.

AVO is also designed to operate across frontier models. Limited experiments paired it with GPT-5.6 Sol on a challenging subset of games. Sol reached matched levels faster in wall-clock time in several cases. Opus used fewer environment actions in matched-level comparisons. The preliminary results suggest complementary operating profiles across models. Broader systematic comparison remains for future work.

These results cover the 25-environment public set using the official scorecard and RHAE metric. They are not results on the semi-private or fully private competition sets.

Where the architecture goes next

NVIDIA AVO research began with autonomous software engineering and high-performance GPU-kernel optimization. ARC-AGI-3 shows the same underlying architecture transfers to a very different reasoning environment. The larger opportunity is building general-purpose agent systems around persistent state, tool use, grounded feedback, recovery, and long-horizon context management. Systems that accumulate evidence and sustain progress across increasingly diverse tasks.

The model matters. The model is not the entire agent.