Scalable causal discovery and reasoning for Agentic AI

Develop scalable frameworks for causal discovery and causal reasoning to provide causal foundations for LLM-based multi-agent Agentic AI systems, enabling robust generalization, safe coordination, and counterfactual or interventional planning under distributional shift.

Background

In surveying foundational gaps for Agentic AI, the paper argues that current systems lack causal foundations, limiting their robustness and capacity to generalize in dynamic environments. This deficiency affects coordination, planning, and error recovery across interacting agents.

The authors explicitly identify scalable causal discovery and reasoning as unsolved challenges, linking them to the need for principled design and reliable multi-agent orchestration beyond ad hoc implementations.

References

Equally critical is the absence of causal foundations as scalable causal discovery and reasoning remain unsolved challenges .

— AI Agents vs. Agentic AI: A Conceptual Taxonomy, Applications and Challenges  (2505.10468 - Sapkota et al., 15 May 2025) in Section: Challenges and Limitations in AI Agents and Agentic AI; Subsubsection: Challenges and Limitations of Agentic AI; Item 8 (Immature Foundations and Research Gaps)

At each stage, identifying an upstream variable would require multiple goodness-of-fit and residual-independence tests. These tests could identify multiple candidates, no candidate, or only inconclusive results, making it unclear which variable should be selected next or whether the procedure should continue. Moreover, because the regressions and tests at later stages depend on earlier ordering decisions, errors or inconclusive outcomes could propagate through the remaining stages.

— Statistical Inference for Bivariate Functional Causal Discovery  (2609.16562 - Prakash et al., 15 Sep 2026) in Section 5, Discussion