Sibyl is an open-source, fully autonomous AI research agent system developed by the Sibyl Research Team and released in May 2026. This system aims to go beyond the capabilities of conventional "paper generators" by autonomously performing the entire research and development (R&D) process, from literature review and hypothesis generation to GPU-based experimentation, result validation, and drafting academic papers, effectively functioning as an AI scientist. By adopting a file-based architecture, the system allows for transparent monitoring of the agent's operational status, memory, and history. Furthermore, its integration with Claude Code provides a foundation for conducting independent machine learning and deep learning research within a powerful multi-agent environment. Similar to how a human researcher improves through writing lab notes and receiving feedback, Sibyl continuously refines itself through file-based experimental records. Previous autonomous AI research agents primarily focused on generating executable workflows or plausible text-based papers. However, they lacked the "scientific judgment" necessary to independently correct and assess situations where experiments fail or data presents inconsistencies. In other words, while they might succeed in producing one-time results, they suffered from the persistent limitation of failing to learn from repeated trial and error, leading to the same mistakes or leaving statistical biases unaddressed. To overcome these limitations, Sibyl introduces a unique self-evolving trial-and-error harness structure. This refers to the core mechanism where the agent does not ignore negative signals, such as errors or duplicate results encountered during experimentation, but instead reduces them into a feedback loop to modify plans or refine validation methods. Just as a Go AI improves its skills by analyzing wins and losses through numerous self-play games, Sibyl converts all data points of success and failure into system behavior guidelines, thereby strengthening the rigor of its research. Researchers in fields such as biotechnology or machine learning can leverage Sibyl to dramatically accelerate idea exploration and tuning tasks, such as developing large-scale drug efficacy prediction models or training neural networks for protein structure analysis. For example, if a researcher provides a general hypothesis for predicting the response of a specific bioactive substance, Sibyl will immediately crawl relevant academic databases, analyze prior research, and design the optimal model architecture and hyperparameter combination. Subsequently, it directly runs the allocated NVIDIA GPU on a virtual machine, recording the learning curve and error signals for each epoch in real-time file format. If gradient vanishing is detected during training, the trial-and-error harness detects the malfunction signal and activates the trial-to-behavior conversion module, automatically modifying the weight initialization technique, thereby performing active system self-correction. Finally, based on the derived optimal model and a quantitative comparison table with the control group, it generates a draft paper in the format of an international conference and provides it to the researcher, allowing the researcher to focus on establishing core ideas instead of simple coding or repetitive experiments.

Sibyl is an open-source, fully autonomous AI research agent system released in May 2026 by the Sibyl Research Team.