Schema is a new agent harness that uses interactive program induction to help LLM agents learn and interact with unknown environments. It raises performance on a benchmark task by 40.5% and solves 100% of a game benchmark.