Self-play has reshaped the landscape of AI, producing agents like AlphaGo and AlphaZero that outmatch human mastery through recursive learning. In this work, we ask: can machines play against themselves not to win games, but to strengthen software? We explore this idea in the realm of software validation, where large language models (LLMs) trained on vast oceans of code hold within them entire worlds of logic, error, and repair. Our proof-of-concept system features three LLM-driven agents: an attacker that crafts flawed code, a defender that hunts for vulnerabilities, and a judge that weighs their wits. In this evolving dance of sabotage and defense, the agents sharpen each other, creating a self-sustaining arms race of refinement. In our preliminary evaluation of adversarial self-play for technical debt detection, we observe improvement in accuracy from 19.73% to 49.66% on a benchmark dataset. Our findings offer an early indication that ideas inspired by self play may contribute to the development of software validation tools.

@inproceedings{paperbot11984,
  title = {Code Wars: Adversarial Self-Play for Evolving Software Validation Tools},
  author = {Merve Astekin and Sagar Sen and Arda Göknil and Hasan Sözer},
  booktitle = {ICSE},
  year = {2026},
  doi = {10.1145/3786582.3786831}
}