$π$-Play: Multi-Agent Self-Play via Privileged Self-Distillation without External Data

$π$-Play addresses a critical bottleneck in the 'Deep Search' agent paradigm (e.g., OpenAI o1, SearchGPT): the sparsity of rewards when agents search for complex information. By treating the 'Question Construction Path' (QCP) as a privileged signal for distillation, it allows agents to learn from their own search process rather than just the final binary outcome. While the 10 forks in 2 days indicate high initial interest from the research community, the project currently lacks a moat beyond the specific algorithmic recipe. The frontier risk is high because OpenAI, Anthropic, and Google DeepMind are aggressively pursuing self-play and synthetic data generation for reasoning models; any breakthrough here is likely to be absorbed into proprietary training pipelines within months. The defensibility is low because the core value is an algorithmic insight rather than a platform, making it easily reproducible by any lab with sufficient compute. It competes with existing paradigms like STaR (Self-Taught Reasoner) and ReST, but focuses specifically on the search-space orchestration.