AI agents performing long-horizon tasks, like migrating CRM data, require more than just internal memory. They depend on a runtime layer, or harness, for execution feedback and state management. Traditionally, developers script agent behavior step-by-step, limiting autonomy and adaptability. EvoHarness-RL, from Meta AI and the University of Illinois Urbana-Champaign, introduces a unified workspace called Belief, Progress, and Experience (BPE) to enhance agent capabilities. Belief tracks the environment, Progress manages subgoals, and Experience stores historical knowledge. Agents interact with BPE using four meta-actions: track, commit, recall, and note. This framework allows agents to learn when and how to access and update their external state. EvoHarness-RL involves supervised fine-tuning to structure data and cost-aware reinforcement learning to optimize tool usage based on computational cost. In benchmark tests, EvoHarness-RL significantly improved the performance of smaller AI models, even matching that of larger, closed-source models. The framework also benefits existing frontier models by enhancing their execution through the BPE prompt-time harness. During training, EvoHarness-RL demonstrates "harness annealing," where agents reduce reliance on external tools for routine tasks, and "harness evolution," where they dynamically adapt strategy based on task complexity. An environment adapter facilitates integration into existing enterprise systems without requiring a complete overhaul of current tools or agent frameworks. EvoHarness-RL represents a shift from scripting agent behavior to creating systems where better behavior can be learned, particularly valuable for long and complex tasks.