Can Environment-Agnostic Agents Solve ARC-AGI-3 from Warped and Noisy Images?

We test an environment-agnostic agent's ability to solve public ARC-AGI-3 games from noisy and geometrically distorted images. The agent receives all information about the board's contents through images, with no textual or numerical grid representation. We apply Gaussian pixel noise and random projective or nonlinear warps, sampled independently for each frame. Geometric distortion makes clicking particularly difficult: the agent must infer coordinates on the original, unwarped board. In additional experiments, we provided cursor feedback showing where proposed board coordinates place a cursor. For GPT-5.6-sol, geometric distortion reduced performance much more than pixel noise alone. Cursor feedback partly restored performance under projective transformations and moderate nonlinear warping. Under the strongest warping and noise tested in the main study, GPT-5.6-sol scored only ∼15% RHAE with or without cursor feedback. By contrast, the same agent with GPT-6-astra achieved a near-ceiling 95.44% RHAE under the same distortions, even without cursor feedback. To probe the limits of GPT-6-astra, we conducted followup experiments without cursor feedback. Adding random board rotation after each action to the strongest main-study setting reduced RHAE only modestly. In separate experiments, extreme pixel noise under strong nonlinear warping substantially reduced performance, but the agent still achieved 17.37% RHAE at the highest noise level tested. Agent-created scripts and denoised, unwarped images found in its workspace suggest that successful play may combine image processing with decisions under substantial uncertainty. Prior exposure to the public games may also have contributed.