Docker is the Bottleneck — Dockerless Fixes AI Coding Agent Training

Training coding agents has a dirty secret: the model isn't the bottleneck — Docker is. Every training repository needs a built image, pinned dependencies, a curated test suite, runners, parsers. A new paper from Shanghai Jiao Tong University and the Douyin Group throws all of that away.

The bottleneck: verification

To train a coding agent you need a reward signal: did the patch fix the bug? The standard answer is "run the tests," which means an executable environment per repo — thousands of Docker images to build and maintain. That infrastructure cost, not model quality, is what caps how much training data agent teams can use.

The idea: judge, don't execute

Dockerless is an environment-free patch verifier. Instead of executing tests, it evaluates a patch the way a senior code reviewer would:

1. Spawn parallel sub-agents that explore the repository, gathering evidence with grep/find — call sites, related tests, conventions
2. Judge the patch with a trained 9B model reading the patch plus the collected evidence

No container, no dependency install, no test run. Static reasoning with an evidence-gathering swarm in front of it.

The results

• 81.0 AUC as a verifier — beating GPT-5.4, GLM-5, and every open-source verifier (+14.3 points)
• Powers a fully environment-free post-training pipeline: SFT data filtering and RL rewards, no execution anywhere
• Models trained this way hit 62.0% SWE-bench Verified, 50.0% Multilingual, 35.2% Pro — matching Docker-based training with zero Docker images

That last line is the headline: the execution environment was supposed to be the ground truth, and a 9B judge matches it.

Why this matters

Verification-by-judgment scales where execution can't: any repo, any language, no infra team. It's also a bet on a broader trend — that trained judges can replace expensive ground-truth harnesses in RL pipelines. If that holds, the moat around big agent-training operations gets a lot shallower.

📄 Paper: arxiv.org/abs/2606.28436

Fun fact from the video: the narration is my own voice, cloned with open-source Qwen3-TTS from a few seconds of reference audio.