programgym-1648

programgym-1648 is a collection of 1,648 Harbor tasks that evaluate an agent's ability to reproduce software from observed behavior. For each task, the agent receives a reference executable and documentation. It must create an original codebase and a compile.sh script that builds a new ./executable with matching behavior. Copying, wrapping, or decompiling the reference executable is not allowed.

Task structure

Each task contains:

Path	Purpose
task.toml	Harbor task metadata and runtime settings.
instruction.md	The task requirements and rules.
environment/	The agent environment and reference executable.
tests/	The verifier and test fixtures.
solution/	Harbor's solution entry point for validation.

Run with Harbor

Requires Docker, a configured model provider, and Harbor 0.20.0 or newer. These tasks use Harbor task schema 1.3 and phase-scoped network policies. Run it by Harbor:

uv tool install 'harbor>=0.20.0'

harbor run \
  -d ProgramGym/programgym-1648 \
  -a claude-code \
  -m your-provider/your-model \
  -n 4

Replace the model ID and provide the corresponding API credentials. The command runs the full dataset with four concurrent trials; use -l 1 for a one-task smoke test. Results are written under jobs/ by default.

Citation

If you find our work helpful, feel free to give us a cite.

@misc{yang2026programgym,
      title={ProgramGym: Scaling Training Environments for Long-Horizon Program Reproduction Agents},
      author={Chengyi Yang and Renxu Bao and Jue Hou and Zongpei Teng and Ziheng Wang and Yujie Lin and Chengsong Huang and Junnan Zhu and Delai Qiu and Yining Wang and Shengping Liu and Fei Long and Jinsong Su},
      year={2026},
      eprint={2610.13395},
      archivePrefix={arXiv},
      primaryClass={cs.SE},
      url={https://arxiv.org/abs/2610.13395},
}