Learning to accomplish tasks in environments with unknown rules remains a fundamental challenge for large language model (LLM) agents. Existing agents typically record their discoveries in prose, which lacks a compact and testable description of how the environment operates. Inspired by how scientists organize observations into testable, predictive theories, we introduce Schema—an agent harness that structures learning and action through interactive program induction.
Within Schema, the LLM agent autonomously decides what to investigate and how to act, expressing its evolving understanding of the environment as executable programs. The harness comprises:
- a persistent program workspace that stores and iteratively refines the environment model;
- a small set of interfaces that allow programs to be checked against interaction history, to plan within the program, and to execute plans under step‑by‑step verification.
This programmatic approach enables the agent to validate hypotheses at each step, rapidly converging on the correct environmental mechanics. Empirically, Schema raises the ARC‑AGI‑3 RHAE score from 58.7% to 99.2% using the same base model, solves 100% of the public DiG‑bench games, and reaches the median performance of the top‑50 human players on MazeBench. Extensive analysis demonstrates Schema’s effectiveness in unknown mechanism discovery, and ablation studies confirm the contribution of each component.
Review