Dockerless: Environment-Free Program Verifier for Coding Agents
arXiv · HuggingFace · ▲103
Abstract (verbatim)
Program verifiers play a central role in training coding agents, including selecting trajectories for supervised fine-tuning (SFT) and providing rewards for reinforcement learning (RL). Standard execution-based verification requires running unit tests inside per-repository environments such as Docker images, incurring substantial environment setup costs. We propose Dockerless, an environment-free agentic patch verifier that evaluates generated code patches without executing them. Rather than simply matching candidate patches to references, Dockerless judges patch correctness using evidence gathered through agentic repository exploration. On a verifier evaluation benchmark, Dockerless outperforms the strongest open-source verifier by 14.3 AUC points. Using Dockerless as both the SFT trajectory filter and the RL reward enables a fully environment-free post-training pipeline. The resulting model reaches 62.0%, 50.0%, and 35.2% resolve rate on SWE-bench Verified, Multilingual, and Pro, respectively. It surpasses the Qwen3.5-9B baseline by 2.4, 8.7, and 2.9 points, matching environment-based post-training.
Background
Background Analysis
1. Technical Context and Real-World Needs
Program verifiers are critical for training automated coding agents (e.g., AI models solving software engineering problems). Their role is to judge whether a code modification correctly resolves an issue, guiding the model’s learning (e.g., filtering high-quality training data or providing reinforcement learning rewards). Traditionally, verification requires executing test cases in isolated environments (e.g., Docker containers per repository), which is essential for real-world applications—whether open-source or enterprise codebases need to ensure AI-generated fixes work as intended.
2. Limitations of Previous Methods
Existing execution-based verification methods face significant challenges. First, environment setup is costly: each repository may require customized dependencies and test scripts, while many real-world scenarios (e.g., private or legacy systems) cannot provide reproducible environments. Second, shallow verification approaches (e.g., comparing code text differences) fail to understand deep code logic, leading to incorrect judgments for complex problems. Even advanced shared environment solutions (e.g., unified Docker images) become performance bottlenecks due to lack of targeted analysis. These limitations make training efficient coding agents expensive and difficult.
3. Proposed Solution
The paper introduces Dockerless, an environment-free program verifier. Its core idea is to let the verifier actively "explore" the codebase: by generating verification questions and dispatching sub-agents to collect contextual evidence (e.g., function call relationships or module dependencies), it ultimately judges the correctness of code modifications. This approach avoids runtime execution, instead leveraging the codebase’s inherent information for reasoning, thus solving environment dependency and shallow analysis issues.
4. Key Differences from Prior Work
Dockerless innovates by shifting verification from "passively matching text" to "actively exploring context." Unlike traditional Docker-dependent methods, it requires no runtime environment; compared to shallow approaches that only compare code differences, it uses multi-agent collaboration to deeply understand code logic. Experiments show this environment-agnostic verification is not only more efficient but also achieves performance comparable to execution-based methods on real tasks, providing a viable path for scaling coding agent training.




