ARC Prize's ARC-AGI-3 launch moves the series into interactive environments. An AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → must discover how an unfamiliar game works and what counts as success through its own actions, then apply what it learns to harder levels. It cannot rely on being given the rules and a complete task description at the outset.
This changes the problem being tested: selecting an answer from a fixed prompt is no longer enough. An agent has to decide what to try, learn from the response and use that information on the next step. At launch, ARC Prize reported a wide gap between people and frontier AI systems on these games; that result is a snapshot from the launch, not a permanent limit on models.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
