AutoResearch is an open-source system that allows AI agents to turn research ideas into experiments, evaluate the results and use evidence to guide the next stage of investigation.
EvoMap has open-sourced AutoResearch, a system designed to help AI agents independently test their own research ideas. Rather than stopping at generating hypotheses or proposing solutions, the platform takes ideas through experimentation and uses the resulting evidence to determine what should happen next.
The system addresses what EvoMap describes as the research verification problem. AI models can generate ideas, write code and analyse results, but a plausible proposal does not necessarily mean that it will work in practice. AutoResearch therefore focuses on testing ideas through measurable experiments instead of relying solely on an AI model’s confidence.
AutoResearch uses multiple AI models to independently generate and cross-review research ideas. Accepted proposals are converted into executable research plans with defined metrics, success criteria, resource budgets and evaluation procedures. Specialised agents then handle different stages, including planning, implementation, experimentation, analysis and review.
The platform also maintains a persistent workspace containing research states, code, experiment logs, metrics, failures and decisions. This allows agents to continue unfinished investigations instead of restarting from scratch. Independent and blind reviews can further challenge conclusions before a research direction is considered complete.
An important part of the approach is that unsuccessful experiments are not simply discarded. Partial results can lead to revised hypotheses and additional experiments, while repeated results that fail to produce meaningful improvements can bring an investigation to a close while preserving what was learned. By making AutoResearch open source, EvoMap aims to provide a framework for AI agents to move from proposing research ideas to systematically testing them through experiments and evidence.
For more information, click here.
