CausalSteward: multi-agent AI copilot for high-dimensional causal discovery

Causal discovery from high-dimensional data remains a major challenge, especially when real-world violations of core assumptions break causal identifiability. To address this, researchers from multiple institutions have introduced CausalSteward (CAST), a novel human-in-the-loop multi-agent system. CAST acts as a copilot that guides users through a divide-conquer-combine pipeline. Large sets of variables are iteratively partitioned into smaller, more manageable clusters using automated tools. Each cluster is then analyzed independently by specialized agents that incorporate retrieval-augmented generation (RAG) to leverage prior knowledge from scientific literature and conditional independence tests for data-driven validation. The results are combined into a coherent global causal model, with the human operator providing oversight and refinement at each step.

This agentic approach offers several advantages. By breaking down high-dimensional problems, CAST reduces computational complexity and improves the reliability of causal conclusions in domains where traditional methods struggle, such as genomics, economics, or climate science. The integration of prior knowledge via RAG helps compensate for limited data, while the human-in-the-loop component ensures transparency and trustworthiness. The authors also use this work to examine both the capabilities and limitations of causal reasoning in multi-agent frameworks. While the system is still experimental, it represents a significant step toward practical, scalable causal inference tools that domain experts can actively collaborate with.