RAT: RunAnyThing via Fully Automated Environment Configuration
RAT achieves 29.6% higher environment setup success rate across all languages.
Automating software engineering tasks at the repository level has long been hindered by the need to manually configure executable environments. A new paper from researchers including Renhong Huang and Dongdong Hua introduces RAT (RunAnyThing), a language-agnostic framework that fully automates this process. RAT works on arbitrary repositories, regardless of programming language, using a multi-stage pipeline that includes semantic initialization to understand the project, a planning mechanism to sequence setup steps, a specialized toolset for actions like installing dependencies, and a robust sandbox for safe execution. This eliminates the labor-intensive manual work that previously limited autonomous code agents.
To rigorously test RAT, the team also created RATBench, a benchmark that reflects the diversity and complexity of real-world repositories. In extensive experiments, RAT achieved state-of-the-art performance, improving the Environment Setup Success Rate (ESSR) by an average of 29.6% over strong baselines. By removing the environment configuration bottleneck, RAT enables code agents to run and test code from any repository autonomously, paving the way for more capable and scalable automated software engineering tools.
- RAT is a language-agnostic framework that automates environment configuration for any code repository.
- It uses a multi-stage pipeline with semantic initialization, planning, specialized tools, and a sandbox.
- On the new RATBench benchmark, RAT improves Environment Setup Success Rate (ESSR) by 29.6% over baselines.
Why It Matters
RAT removes a key bottleneck for autonomous code agents, enabling them to run code from any repository automatically.