MEnvAgent: Scalable Polyglot Environment Construction for Verifiable Software Engineering

Researchers introduce MEnvAgent, a framework for automatically constructing executable software environments across multiple programming languages, addressing a critical bottleneck in LLM agent training. The system generates verifiable datasets and reduces computational costs by 43%, enabling the creation of MEnvData-SWE, the largest open-source polyglot dataset of Docker environments for software engineering tasks.

MEnvAgent tackles a fundamental infrastructure problem limiting LLM agent development in software engineering. Creating verifiable, executable environments across diverse programming languages requires significant engineering overhead—developers must provision, test, and validate complex setups spanning Python, JavaScript, Java, Go, Rust, and other languages. This bottleneck has constrained the availability of high-quality training datasets, directly limiting how well AI agents can learn to solve real software engineering problems.

The framework's Planning-Execution-Verification architecture autonomously resolves construction failures, while its Environment Reuse Mechanism patches existing setups incrementally rather than rebuilding from scratch. These design choices address the core inefficiency: traditional approaches rebuild entire environments for each task variation, consuming substantial computational resources. By improving Fail-to-Pass rates by 8.6% while cutting time costs by 43%, MEnvAgent makes large-scale dataset generation economically feasible.

For the AI research community, this unlocks substantial value. MEnvData-SWE provides 1,000 realistic, verifiable task instances—a resource developers can use to train and benchmark LLM agents at scale. This democratizes access to quality training data, potentially accelerating progress in AI-assisted software engineering across diverse programming contexts.

Looking forward, the framework's modular design suggests it could extend beyond the 10 languages currently supported. As LLM agents become increasingly deployed in production engineering workflows, reliable, language-agnostic environment construction becomes a critical infrastructure component. The open-source release signals momentum toward standardized tooling for this previously ad-hoc process.

• →MEnvAgent reduces computational overhead by 43% through incremental environment patching instead of full reconstruction
• →MEnvData-SWE provides the largest open-source polyglot dataset of verifiable Docker environments for software engineering tasks
• →The framework improves Fail-to-Pass rates by 8.6% across 1,000 tasks spanning 10 programming languages
• →Autonomous resolution of construction failures enables scalable generation of verifiable task instances for LLM agent training
• →Open-source release democratizes access to infrastructure for building and benchmarking AI-powered software engineering tools