Dockerless verifier boosts coding agents without environment setup
Skip Docker entirely: new verifier matches code patches via repo exploration, not execution.
Training coding agents traditionally requires executing generated code patches inside isolated environments like Docker containers to verify correctness. This process is slow, costly, and complex to scale. In a new paper, researchers from multiple institutions introduce Dockerless, a verifier that evaluates patches without any runtime execution. Instead of running unit tests, Dockerless uses an agentic approach: it explores the target repository, collects contextual evidence from code, documentation, and test files, then judges whether a patch is correct based on that evidence.
On a dedicated verifier evaluation benchmark, Dockerless beats the strongest open-source verifier by 14.3 AUC points. When integrated as both a trajectory filter for supervised fine-tuning and as a reward model for reinforcement learning, it enables a fully environment-free post-training pipeline. The resulting trained model achieves 62.0% on SWE-bench Verified, 50.0% on Multilingual, and 35.2% on Pro, surpassing a Qwen3.5-9B baseline by 2.4, 8.7, and 2.9 points respectively—matching the performance of environment-based post-training without the infrastructure overhead.
- Dockerless eliminates the need for Docker containers or runtime execution for code verification.
- It uses agentic repository exploration to judge patch correctness, outperforming open-source verifiers by 14.3 AUC points.
- Enables a fully environment-free post-training pipeline, achieving 62.0% on SWE-bench Verified vs 59.6% for Qwen3.5-9B baseline.
Why It Matters
Removes a major infrastructure bottleneck in coding agent training, making high-quality verification accessible without heavy compute.