VIGA approaches vision as inverse-graphics through a dual-role agent that alternates between generating scene programs and verifying rendered output across multiple viewpoints. The agent writes and executes code using tools for planning, scene queries, and asset retrieval, while maintaining contextual memory of plans, code changes, and render history. This write-run-compare-revise loop is self-correcting and operates without finetuning, making it a general framework for programmatic visual reconstruction tasks.