2025 marks a significant shift in AI, particularly with the emergence of agents capable of performing complex tasks in realistic settings, rather than just responding in chat interfaces. The focus has moved toward evaluating how well AI models handle multi-step tasks using tools. Surge HQ tested nine AI models on 150 tasks within a reinforcement learning (RL) environment, revealing that while GPT-5 and Claude Sonnet 4.5 excelled, even they struggled with over 40% of the tasks. This highlights the ongoing challenge of developing agents that can operate effectively in real-world scenarios. To build these RL environments, three critical components are necessary: a coherent world model, a set of entities, and a tool system. For example, Corecraft, Inc. serves as a simulated online retailer, where AI models act as customer support agents. The tasks range from simple queries, like refund counts, to complex problem-solving, such as resolving product compatibility issues. This role helps gauge the economic potential of AI in everyday applications, emphasizing the need for agents to master foundational skills before they can tackle more complex challenges. The Hierarchy of Agentic Capabilities outlines the skills AI models need. Starting with basic tool use and goal formation, models must progress to higher-order skills like adaptability and common-sense reasoning. Current models, such as GPT-4o, Mistral Medium, and Nova 1 Pro, often fail at the foundational level, struggling with basic task breakdown and execution. Specific errors included misidentifying inputs and failing to follow established protocols for multi-step tasks, underscoring the necessity for improvement in these areas before models can function reliably as agents in diverse environments.