The Autoresearcher Intern: Evidence-Gated Human-AI Collaboration for Scientific Workflow Automation

We present the autoresearcher intern, a repository-centered pipeline for human-guided scientific prototyping with language-model agents. The system decomposes research into bounded planning, execution, review, summarization, and checkpoint-supervision roles; stores plans, commands, metrics, artifacts, reviews, and claim boundaries as auditable repository state; and escalates high-stakes decisions to a stronger model such as gpt-5.5-pro. Across two live reinforcement-learning case studies, the pipeline produced scoped positive evidence, decision-relevant negative results, and stopping decisions that prevented unsupported scaling. We argue that the autoresearcher intern is a practical collaboration paradigm for fast, large-scale prototyping of scientific ideas.