Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Danlong Yuan, Wei Wu, Zhengren Wang, Xueliang Zhao, Huishuai Zhang, Dongyan Zhao

TL;DR

SWE-MiniSandbox addresses the scalability and accessibility limitations of container-based reinforcement learning for software engineering agents by introducing a container-free sandbox that uses per-instance kernel-level isolation and a light, reusable environment pre-caching pipeline. The approach significantly reduces storage and setup overhead while maintaining training performance, enabling multi-node RL at resource-constrained scales. Key contributions include a practical container-free isolation mechanism, an efficient I/O-aware pre-caching strategy with a disciplined per-task I/O budget, and seamless integration with existing SWE tooling (SWE-Rex, SWE-agent, SkyRL) for distributed RL. The results demonstrate substantial storage and time savings (roughly 5% of container-based baselines and ~25% of setup time) with comparable evaluation fidelity, highlighting the framework’s potential to democratize large-scale SWE-agent experimentation.

Abstract

Reinforcement learning (RL) has become a key paradigm for training software engineering (SWE) agents, but existing pipelines typically rely on per-task containers for isolation. At scale, pre-built container images incur substantial storage overhead, slow environment setup, and require container-management privileges. We propose SWE-MiniSandbox, a lightweight, container-free method that enables scalable RL training of SWE agents without sacrificing isolation. Instead of relying on per-instance containers, SWE-MiniSandbox executes each task in an isolated workspace backed by kernel-level mechanisms, substantially reducing system overhead. It leverages lightweight environment pre-caching techniques to eliminate the need for bulky container images. As a result, our approach lowers disk usage to approximately 5\% of that required by container-based pipelines and reduces environment preparation time to about 25\% of the container baseline. Empirical results demonstrate that SWE-MiniSandbox achieves evaluation performance comparable to standard container-based pipelines. By removing the dependency on heavy container infrastructure, SWE-MiniSandbox offers a practical and accessible foundation for scaling RL-based SWE agents, particularly in resource-constrained research environments.

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

TL;DR

SWE-MiniSandbox addresses the scalability and accessibility limitations of container-based reinforcement learning for software engineering agents by introducing a container-free sandbox that uses per-instance kernel-level isolation and a light, reusable environment pre-caching pipeline. The approach significantly reduces storage and setup overhead while maintaining training performance, enabling multi-node RL at resource-constrained scales. Key contributions include a practical container-free isolation mechanism, an efficient I/O-aware pre-caching strategy with a disciplined per-task I/O budget, and seamless integration with existing SWE tooling (SWE-Rex, SWE-agent, SkyRL) for distributed RL. The results demonstrate substantial storage and time savings (roughly 5% of container-based baselines and ~25% of setup time) with comparable evaluation fidelity, highlighting the framework’s potential to democratize large-scale SWE-agent experimentation.

Abstract

Reinforcement learning (RL) has become a key paradigm for training software engineering (SWE) agents, but existing pipelines typically rely on per-task containers for isolation. At scale, pre-built container images incur substantial storage overhead, slow environment setup, and require container-management privileges. We propose SWE-MiniSandbox, a lightweight, container-free method that enables scalable RL training of SWE agents without sacrificing isolation. Instead of relying on per-instance containers, SWE-MiniSandbox executes each task in an isolated workspace backed by kernel-level mechanisms, substantially reducing system overhead. It leverages lightweight environment pre-caching techniques to eliminate the need for bulky container images. As a result, our approach lowers disk usage to approximately 5\% of that required by container-based pipelines and reduces environment preparation time to about 25\% of the container baseline. Empirical results demonstrate that SWE-MiniSandbox achieves evaluation performance comparable to standard container-based pipelines. By removing the dependency on heavy container infrastructure, SWE-MiniSandbox offers a practical and accessible foundation for scaling RL-based SWE agents, particularly in resource-constrained research environments.
Paper Structure (27 sections, 3 equations, 11 figures, 8 tables)

This paper contains 27 sections, 3 equations, 11 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. SWE Agent Framework
  4. Efforts to Scale SWE Environments
  5. Preliminaries
  6. Method
  7. MiniSandbox Launch and Isolation
  8. MiniSandbox Pre-Caching
  9. Pre-Caching Pipeline
  10. I/O Bottleneck
  11. Per-task I/O budget model.
  12. Caching Stage of SWE-bench and SWE-smith
  13. RL Integration and Distributed Training Behavior
  14. Experiments
  15. Experimental Setup
  16. ...and 12 more sections

Figures (11)

  • Figure 1: Agent Isolation Strategies: Contrasting our per-instance, namespace-based MiniSandbox (left) with conventional container-based isolation (right).
  • Figure 2: Environment Pre-Caching Pipeline: The workflow for building and archiving reusable task environments.
  • Figure 3: Rollout time comparison between SWE-MiniSandbox and a container-based framework in the 3B-RL setting (step 50).
  • Figure 4: Breakdown of environment preparation time components for SWE-MiniSandbox in the 3B-RL setting.
  • Figure 5: The test output from case psf__requests-2317
  • ...and 6 more figures