ToolBrain: A Flexible Reinforcement Learning Framework for Agentic Tools

Training language-enabled agents to reliably use tools remains a significant challenge, often hindered by complex frameworks and high computational costs. We present ToolBrain, an open-source platform that addresses this challenge through the Coach–Athlete paradigm, an architectural abstraction that simplifies the application of reinforcement learning (RL) to agentic workflows. We evaluate ToolBrain by adapting a compact language model to three representative tasks: multi-step information retrieval, quantitative reasoning, and real-world API interaction. Our results demonstrate that ToolBrain substantially improves the performance of compact models, enabling them to solve complex tool-using tasks efficiently. The accompanying demonstration video can be viewed at https://youtu.be/FIgfg-y0sXw