ZeroPaper is an autonomous system that takes a research domain as input and produces a publication-candidate manuscript as output, with no human intervention between launch and completion. The system runs ten numbered stages and six adversarial gates, coordinates roughly thirty specialized agents, and produces finished LaTeX for economics at roughly two dollars per paper amortized under a flat-fee Claude Code Max subscription (approximately zero marginal cost per additional paper).
From ten premises about how LLMs behave and what they cost, we derive six design principles for long-running autonomous systems: keep state and control flow outside the model, spend context carefully, delegate aggressively, verify rather than trust, make termination mechanical, and run independent work in parallel. ZeroPaper runs on Claude Code, OpenAI Codex, and Gemini CLI, with variants for finance and macro and extensions for empirical analysis and LLM experiments.