A team of computer scientists and biologists has released CellForge, an open-source system that uses a crew of specialized large-language-model agents to design, code and run computational models of virtual cells with no human intervention. The paper, posted on arXiv on 4 August, shows the platform converting raw single-cell multi-omics data and a high-level research objective directly into an optimized model architecture and executable code. CellForge assigns role-specific agents—such as Task Analysis, CPU and GPU coding, and rendering helpers—to collaborate until they reach consensus on a modelling strategy. In tests across six datasets covering gene knockouts, drug treatments and cytokine stimulations, the approach delivered models that were on average 22% more accurate than task-specific state-of-the-art baselines, according to the authors. Backers say the technology could shorten the virtual-cell research cycle and accelerate drug-discovery workflows by automating programming tasks that typically require cross-disciplinary teams. The full codebase has been made available on GitHub for academic and commercial use.