SWE-MiniSandbox: Container-Free Reinforcement Learning for Building Software Engineering Agents

The SWE-MiniSandbox paper presents a novel container-free reinforcement learning environment designed for developing and testing software engineering agents. This sandbox facilitates efficient training and evaluation of AI agents capable of automating various software development tasks, from code generation to bug fixing, without the overhead of containerization.