An agent environment defines the interaction interface, state transitions, and feedback loop through which a large language model (LLM) learns agentic behavior via multi-turn reinforcement learning (RL). Unlike supervised fine-tuning on fixed, off-policy trajectories, environment-based RL allows agents to learn from their own interactions and outcomes. Recent advances in RL with verifiable rewards (RLVR) have shown that this paradigm can push agents beyond the capabilities captured by demonstration data and improve generalization to novel tasks. However, LLM-based agent training remains an emerging field, and scaling it requires more than simply moving from single-turn to multi-turn rollouts: it requires reliable large-scale interaction through concurrent tool execution, sandboxing, state management, and efficient rollout generation, as well as diverse environments with trustworthy verifiers and reward signals. These challenges motivate research on environment synthesis, automatic curricula, and self-evolving environments, while raising several central questions: How can environments scale across domains such as science, coding, office work, terminal and GUI interaction, embodied control, and multi-agent collaboration, while also supporting increasingly complex and long-horizon tasks? How can tasks, rewards, and verifiers be generated and adapted automatically? And how can reliable supervision be maintained while avoiding reward hacking, spurious shortcuts, and brittle evaluation protocols?