SkillGym: Training Skill-Use Agents with Automatic Verifiable Environment Generation

SkillGym. Training Skill-Use Agents with Automatic Verifiable Environment Generation. By Renxi Wang and four others, from the Mohamed bin Zayed University of Artificial Intelligence. Published on the twenty-ninth of September. Its code repository is a placeholder: a readme, and nothing to run yet.

Skills give LLM agents professional knowledge and guidance for long-horizon, complex tasks, and recent agent paradigms and harnesses have taken them up widely. Two questions have had little attention, though. Where does reliable training data for skill use come from? And how should an agent actually be trained to use skills?

SkillGym answers both with one automatic pipeline. It builds verifiable environments, collects trajectories inside them, and trains agents that use skills. The data problem and the training problem are handled together: the environments produce the trajectories, and the trajectories become the training set. Verification is the thread through all of it, since only trajectories that pass a check are kept.

The pipeline starts by crawling a large volume of skills from the internet and keeping only those whose workflows run reproducibly offline. A skill that cannot be rerun reliably cannot anchor a task that has to be checked. A builder-reviewer pipeline then constructs tasks with controlled difficulty, across four task types. Every task comes with a reference solution and an executable verifier, which is what makes the environments verifiable. Using this pipeline, the authors build six point eight thousand environments and collect nineteen thousand verified successful trajectories for supervised finetuning.

Finetuning on those trajectories improves LLMs from different families and of different sizes, from two billion up to one hundred and twenty-two billion parameters, across four skill-use benchmarks. On two of those benchmarks, a Qwen3.5-9B model trained this way outperforms the untrained three hundred and ninety-seven billion parameter model. A model of nine billion parameters, after this finetuning, beats a vastly larger model that has not had it.

Looking closer, training teaches agents to invoke skills at all: the rate at which an agent reads the relevant skill rises from twenty-eight percent to ninety-six percent. The gains also hold across reasoning structures. They extend to task types that make up only a minority of the training data, and to skills held out from training entirely.

The jump from twenty-eight to ninety-six percent on simply opening the right skill shows what the training changed: agents learned to open the skill in the first place. SkillGym gets there through data: thousands of environments that can check their own answers, and trajectories that pass the check.