DeepSeek researcher's open-source AI agent writes papers, runs 285B-parameter experiments

DeepSeek senior researcher Chen Deli has open-sourced a personal project called Deli AutoResearch SKILL — and along with it, the fourth academic survey paper written entirely by an AI agent.

The project is a single protocol file, SKILL.md, with no executable code. It defines how to manage state persistence over long research tasks, prevent infinite loops, and use a heartbeat watchdog to keep the agent on track. The protocol also enables the agent to spawn sub-agents and simulate multiple roles, automating the entire research workflow.

The accompanying paper, a self-play review, highlights a major breakthrough: for the first time, an AI agent autonomously planned a GPU experiment, ran reinforcement learning on DeepSeek's 285B-parameter model using the GRPO algorithm, and completed the full research loop — from experiment design and coding to debugging and drawing conclusions. In a simulated peer review, it scored 8.6 out of 10.

Chen had previously used the same approach to produce three other survey papers. The first one went through about 60 agent iterations over roughly 10 hours, evolving from version V1 to V5. The entire workflow not only slashes the manual overhead of long-cycle research but also validates the feasibility of an AI-native research path.