ByteDance researchers have introduced Dockerless, an environment-free patch verifier designed to evaluate code generated by AI coding agents without requiring execution-based verification. Standard methods typically rely on running unit tests within isolated environments like Docker containers, which involves high setup overhead. Instead, Dockerless assesses the correctness of code patches by gathering evidence through agentic repository exploration. This approach enables developers to perform post-training tasks, such as filtering trajectories for supervised fine-tuning and calculating rewards for reinforcement learning, without the associated costs of maintaining repository-specific environments.

On a specialized verifier evaluation benchmark, Dockerless surpassed the performance of the strongest available open-source verifier by 14.3 AUC points. By replacing traditional execution environments with this verification method, the team achieved a fully environment-free post-training pipeline. The resulting model demonstrated significant performance gains on the SWE-bench benchmark suite, attaining a resolve rate of 62.0% on Verified, 50.0% on Multilingual, and 35.2% on Pro tasks. These results represent an improvement over the Qwen3.5-9B baseline by 2.4, 8.7, and 2.9 percentage points, respectively, effectively matching the capabilities of systems that rely on environment-based post-training.

ByteDance researchers have introduced Dockerless, an environment-free patch verifier designed to evaluate code generated by AI coding agents without requiring execution-based verification. Standard methods typically rely on running unit tests within isolated environments like Docker containers, which involves high setup overhead. Instead, Dockerless assesses the correctness of code patches by gathering evidence through agentic repository exploration. This approach enables developers to perform post-training tasks, such as filtering trajectories for supervised fine-tuning and calculating rewards for reinforcement learning, without the associated costs of maintaining repository-specific environments.

On a specialized verifier evaluation benchmark, Dockerless surpassed the performance of the strongest available open-source verifier by 14.3 AUC points. By replacing traditional execution environments with this verification method, the team achieved a fully environment-free post-training pipeline. The resulting model demonstrated significant performance gains on the SWE-bench benchmark suite, attaining a resolve rate of 62.0% on Verified, 50.0% on Multilingual, and 35.2% on Pro tasks. These results represent an improvement over the Qwen3.5-9B baseline by 2.4, 8.7, and 2.9 percentage points, respectively, effectively matching the capabilities of systems that rely on environment-based post-training.