The AI research closed-loop has been successfully run, with zero human intervention and self-play—this wave is the agent competing against itself.

Crypto.com reported that DeepSeek senior researcher Chen Deli open-sourced Deli AutoResearch, with the fourth review written entirely by an autonomous agent. AI calls sub-intelligent agents to collaborate and complete the entire research process through the skill.md protocol file. The self-play review independently planned GPU experiments for the first time without human intervention, and conducted reinforcement learning training on the 285B parameter DeepSeek model, using the GRPO algorithm to complete the research loop from design to conclusion, achieving an 8.6/10 in simulated peer review. Previously, this method produced three reviews, with the first approximately 60 iterations and a total time of about 10 hours, validating the feasibility of AI-native research pathways.