Causal Inference Frameworks for Trustworthy Decision-Making in Autonomous Agents

Abstract

Autonomous agents that only predict, without reasoning about intervention and counterfactual outcomes, cannot reliably
answer the question every acting system implicitly faces: what would happen if I did this instead. This article synthesizes the
literature on causal inference frameworks as a foundation for trustworthy decision-making in autonomous agents, tracing the
path from foundational estimation tools such as propensity scores, marginal structural models, and invariant prediction,
through causal discovery methods that learn structure from data rather than assume it, to causal reinforcement learning
architectures that fold counterfactual reasoning directly into an agent's policy loop. We give particular weight to results with
concrete, verifiable numbers: counterfactual credit assignment lifting multi-agent win rates well above non-causal baselines, a
targeted causal intervention recovering an imitation-learning agent from a spurious shortcut in a fraction of the queries a
standard baseline needs, and a human study showing that causal explanations measurably improve a person's ability to
predict an autonomous agent's next action. Read together, the evidence supports causal inference as a genuine, increasingly
well-evidenced foundation for trustworthy autonomous decision-making, while leaving formal safety guarantees, causal
discovery at real deployment scale, and the link between causal explanation and human trust as a substantially open research
agenda