Language‑model agents for data analysis usually need many annotated tasks to learn effective interaction with executable environments such as databases. High‑quality annotations demand heavy manual effort, yet an executable environment already supplies rich interaction structure and verifiable outcomes. In this sense, the environment itself acts as a free lunch, offering a set of verifiable tasks without human annotation.\ \ We introduce EnvLunch, a warm‑up stage that converts this environmental signal into agent training before downstream reinforcement learning, without any human‑written or model‑generated task descriptions. EnvLunch works by: (1) sampling valid action trajectories; (2) executing them to obtain the final observation; (3) training the agent to reconstruct a functionally equivalent trajectory given only that final observation. This label‑free warm‑up improves performance by over 9 points on average across all downstream training budgets; with few labels, the baseline requires 8–32× more annotated examples to reach the accuracy that EnvLunch attains at its smallest budget.\ \ Defining tasks solely by the environment yields larger gains than letting the model generate its own tasks on the same environments, and the warm‑up also boosts fine‑tuning on benchmarks beyond data analysis. Thus, EnvLunch saves annotated tasks and provides a stronger starting point for fine‑tuning agentic models when labeled data are scarce.\ \ Review