Formalizes task-agnostic environment preprocessing, where an LLM-agent “studying system” explores a new environment under a budget and builds reusable artifacts (indices, scripts, procedural guidance) before the downstream task is known. Compares unaided and archive-equipped meta-agents against fixed synthetic-practice and corpus-processing baselines across six heterogeneous benchmarks; a meta-agent variant achieves the best average reward on five of six. Larger study budgets do not reliably improve downstream reward, but studied artifacts reduce the test-time sampling needed to reach a given score.
