An AI agent now checks its own work by watching its code actually run, not just eyeballing the final picture.
Researchers built VETR, short for Verified Execution Trajectories for Refinement, to help AI agents get better at writing and fixing code that builds 3D scenes in Blender. Older systems just guess at new code based on how the rendered image looks, with no record of what the program actually did. VETR instead tracks which edits took effect, vanished, or clashed when the code ran, then uses that execution record to generate a constrained set of candidate fixes. Those fixes are checked against the captured runtime state before a separate visual verifier judges whether the result actually matches the goal.
Final-image-only feedback is a known weak point for agents that write code to generate graphics: two renders can look alike for completely different underlying reasons, and vague picture-based corrections can propagate the wrong fix. By grounding corrections in what the program actually executed, VETR beat prior refinement methods BlenderAlchemy and VIGA on 280 BlenderGym tasks and 27 BlenderBench tasks, using both GPT-5.6-luna and Qwen3.8-27B as the underlying models.
It is still a lab result on a narrow slice of work - editing Blender scenes through code - not a general fix for AI coding agents, and the comparison is against academic baselines, not shipped tools.